Intelligent techniques and optimization algorithms in textile colour management: a systematic review of applications and prediction accuracy

Abstract

Based on a selection of 101 articles published from 2013 to 2022, this study systematically reviews the application of intelligent techniques and optimization algorithms in textile colour management. Specifically, the study explores how these techniques have been applied to four subfields within textile colour management: colour matching and prediction, colour difference detection and assessment, colour recognition and segmentation, and dye solution concentration and decolourization. Following an introduction to intelligent techniques and optimization algorithms in textile colour management, the study describes the specific applications of these techniques in the field over the past decade. Descriptive statistics are used to analyse trends in the use of these techniques and optimization algorithms, and comparative performances indicate the effectiveness of the techniques and algorithms. The study finds that the primary intelligent techniques used in the field of textile colour management include artificial neural networks (ANN), support vector machines (SVM) such as SVM, LSSVM, LSSVR, SLSSVR, FWSVR, fuzzy logic (FL) and adaptive neuro-fuzzy inference systems (ANFIS), clustering algorithms (e.g., K-means, FCM, X-means algorithms), and extreme learning machines (ELM) such as ELM, OSLEM, KELM, RELM. The main optimization algorithms used include response surface methodology (RSM), genetic algorithms (GA), particle swarm optimization (PSO), and differential evolution (DE). Finally, the study proposes a comparison of the performance of intelligent techniques and optimization algorithms, summarizes the relevant research trends, and suggests future research opportunities and directions, besides stating the limitations of this paper.

Introduction

To enhance prediction accuracy, intelligent techniques and optimization algorithms are employed to identify and classify patterns in data. Artificial Neural Networks (ANN), K-means clustering algorithms, and Adaptive Neuro-fuzzy Inference Systems (ANFIS) are among the commonly used intelligent techniques (Almodarresi et al., 2013; Pan et al., 2013; Vasseghian & Dragoi, 2018). Likewise, a few examples of commonly used optimization algorithms are Ant Colony Optimization (ACO), Particle Swarm Optimization (PSO) algorithm, and Genetic Algorithm (GA) (Aryafar et al., 2019; Chaouch et al., 2019a; Zhang & Yang, 2014). These techniques and algorithms can serve as effective alternative tools for identification, classification, and prediction in the domain of textile colour management (He et al., 2021; Liu et al., 2022).

In the field of textile colour management, there are numerous problems that need to be addressed, including colour matching and prediction, colour difference detection and assessment, colour recognition and segmentation, as well as dye solution concentration and decolourisation. These issues are complex and cannot be easily solved by simple linear regression models. However, application of intelligent techniques and optimisation algorithms can help enhance the efficiency of optimisation, identification, and prediction. This, in turn, reduces the number of repeated experiments, leading to better solutions to these problems. For instance, ANN is a type of intelligent technique that can make predictions without requiring a mathematical description of the phenomena involved in the process, thus requiring less time to develop models (Khataee et al., 2013). K-means clustering methods can analyse the colour segmentation of printed fabrics in different colour spaces by minimising the sum of the squared distances between all points and the cluster centroids (Pan et al., 2013). ANFIS can combine the learning capabilities of neural networks with the capabilities of fuzzy logic to generate improved predictions (Haji & Payvandy, 2020). Optimisation algorithms, such as ACO, are computationally efficient methods for obtaining good solutions to combinatorial optimisation problems (Chaouch et al., 2019a). Meanwhile, the PSO algorithm is capable of continuously searching the problem space by representing random answers in the search and optimal solution region (Aryafar et al., 2019). Additionally, a combination of neural networks can be used to find the optimal weights and biases, thereby improving the performance of artificial neural networks (Aryafar et al., 2019). Alternatively, GA is a search heuristic in artificial intelligence that mimics natural selection (Zhang & Yang, 2014).

In the field of textile colour management, the desired colour is achieved through dyeing, which involves addressing the challenges of colour matching and prediction. Specifically, an appropriate colour formulation needs to be developed to get the same colour as the given reference colour by predicting the dyes and their respective concentrations (Chaouch et al., 2022). Traditional methods typically rely on experience of practitioners or the Kubelka–Munk theorem, but the accuracy of colour matching is hindered by various factors such as fibres, yarns and finishing processes (Şahin et al., 2022). Thus, investigating the problem of colour matching and prediction through intelligent techniques and optimization algorithms is essential. Moreover, after dyeing, the quality of the fabric's colour serves as a significant indicator for product testing (Zhang & Zhou, 2022). Traditional colour classification methods involve manual classification by experienced and skilled workers, who are susceptible to visual fatigue and subjectivity (Zhang & Zhou, 2022). Therefore, intelligent techniques and optimization algorithms-based colour difference detection and assessment models have been developed to meet the requirements of automated production of dyed fabrics. Furthermore, with diversification of colours in clothing, the demand for printed fabrics with complex colour patterns has increased (Qian et al., 2022). In the production of fabrics, colour recognition and segmentation are indispensable steps (Qian et al., 2022). However, the traditional methods of colour recognition and segmentation are not only labour-intensive but also unstable (Qian et al., 2022). Therefore, research on colour recognition and segmentation of multicoloured fabrics based on intelligent techniques and optimization algorithms has been initiated. Finally, concerning the concentration and decolourization of dye solutions in textile wastewater, a combination of intelligent techniques and optimization algorithms can effectively optimize the modelling process, save time and experimental costs, and improve the efficiency of pollutants removal (Kothari et al., 2022).

However, there is a dearth of comprehensive and systematic reviews, in extant literature, of applications of intelligent techniques and optimization algorithms in the field of textile colour management. While previous reviews of applications of intelligent techniques have been scattered and focused on textile wastewater decolourization, textile supply chain, and textile manufacturing processes (He et al., 2021; Liu et al., 2022; Ngai et al., 2014), a comprehensive review of the existing literature is critical to inform applications of intelligent techniques and optimization algorithms in the field of textile colour management. This review focuses on the last decade, i.e., from 2013 to 2022, and considers research articles published in journals or conference articles on applications of intelligent technologies and optimization algorithms in textile colour management. The research articles were selected using an advanced search in Web of Science with the following keywords: "TS = (textile OR fabric) AND TS = (colour fading OR colour prediction OR colour management OR colour recipe OR fading) AND TS = (ANN OR neural network OR intelligent techniques OR algorithm OR intelligent)." After removing articles that were not relevant to the topic, we selected 101 articles. For classifying different issues related to the field of textile colour management, a subjective approach was employed. This approach relied mainly on the authors' judgment by designing sub-category headings and classifying each article according to their own opinions (Liu et al., 2022).

This paper presents a systematic review of previously published literature on applications of intelligent techniques and optimization algorithms in the field of textile colour management. The objectives of this review are as follows:

• 1. To describe specific applications of intelligent techniques and optimization algorithms in the field of textile colour management.

• 2. To analyse the current trends in the use of intelligent techniques and optimization algorithms in the field of textile colour management.

• 3. To summarize the relevant research trends and identify future research opportunities for application of intelligent techniques and optimization algorithms in the field of textile colour management.

Methods

Specific applications of intelligent techniques and optimisation algorithms in the field of textile colour management

After reviewing 101 research articles published in the last decade (2013–2022) in the field of textile colour management, a subjective approach was used to classify these articles into the following four categories: (1) colour matching and colour prediction; (2) colour difference detection and assessment; (3) colour recognition and colour segmentation; and (4) dye solution concentration and decolourisation. Specifically, we have identified specific applications of intelligent techniques and optimization algorithms in this field and categorized them as follows: colour matching and colour prediction (34 research articles); colour difference detection and assessment (15 research articles); colour recognition and colour segmentation (20 research articles); and dye solution concentration and decolourisation (32 research articles).

Colour matching and colour prediction

For colour matching and prediction, a variety of intelligent techniques and optimization algorithms are available. These include Artificial Neural Networks (ANN) and Recurrent Neural Networks (RNN), Fuzzy Logic (FL) and Adaptive Neuro-Fuzzy Inference Systems (ANFIS), Response Surface Methods (RSM) and Taguchi methods, Combinatorial Optimization (CO), Ant Colony Optimization (ACO), Genetic Algorithms (GA), Grey Wolf Optimization (GWO), PSO-FMINCON algorithms (PSO-FMIN), Multi-Objective Evolutionary Algorithms (MOEA), and Non-dominated Sorting Genetic Algorithm II (NSGAII). A total of 34 research articles related to colour matching and prediction published in the past decade have been compiled and are presented in Table 1, sorted by algorithms. Besides, a meticulous exploration of the technical nuances and applications of these methodologies will ensue, with the objective of elucidating their respective merits and limitations.

Table 1 Specific applications of intelligent techniques and optimisation algorithms in the field of colour matching and colour prediction

Firstly, ANNs have been substantiated as potent tools for conducting colour predictions and colourant prognostications (Almodarresi et al., 2013; Hasanzadeh et al., 2013). Specifically, ANNs can be trained to emulate the behaviours of professional colourists, thereby learning the relationships between colourants and specific colour coordinates (Almodarresi et al., 2013). Nonetheless, conventional ANNs may encounter difficulties in processing sequential data. To mitigate this issue, RNNs have been introduced (Zhang et al., 2021). RNNs permit data to be input in a sequential format and share weight matrices (Zhang et al., 2021). However, RNNs are not devoid of issues, such as determining the number of network layers (Zhang et al., 2021). To ameliorate these issues, researchers have introduced dense blocks and employed Long Short-Term Memory (LSTM) blocks to preclude information loss (Zhang et al., 2021). Both techniques have significantly enhanced the performance of RNNs.

In addition, fuzzy logic, emanating from the foundational fuzzy set theory, navigates through four pivotal components: the fuzzification interface, the knowledge base, decision logic, and defuzzification (Hossain et al., 2015). It mitigates uncertainties and imprecisions by employing fuzzy sets and operational rules of fuzzy mathematics. Nonetheless, selecting a judicious number of membership functions and parameters to safeguard system accuracy whilst circumventing an escalation in complexity emerges as a formidable challenge (Hossain et al., 2015). In an extended discourse, the ANFIS model amalgamates concepts from fuzzy logic and neural networks, endeavouring to optimize the parameters of the fuzzy logic system through the exploitation of the learning capability intrinsic to neural networks (Haji & Payvandy, 2020).

Besides, the RSM proffers a methodological approach to ascertain the multivariate regression equation between the response variable and a set of design variables, facilitated through quantitative data (Hasanzadeh et al., 2013). In a parallel vein, the Taguchi method also furnishes a methodology to discern the optimal parameter combination by conducting a minimized quantity of experiments (Hossain et al., 2016a, 2016b). However, the Taguchi method encounters challenges, notably an incapacity to invariably determine the impact of interactions and a deficiency in ensuring the predictive accuracy of the model (Hossain et al., 2016a, 2016b).

Concomitantly, metaheuristic methodologies, exemplified by ACO and GWO, proffer mechanisms for procuring considerably propitious solutions to stringent Combinatorial Optimization (CO) dilemmas within an acceptable computational temporal framework (Chaouch et al., 2019a; Vadood & Haji, 2022a). The PSO-FMINCON amalgamated algorithm assimilates the merits of Particle Swarm Optimization and FMINCON, furnishing a more exhaustive optimization stratagem (Vadood & Haji, 2022a). Notwithstanding the incapacity of these algorithms to invariably assure the identification of the global optimum, their efficacy has been substantiated in an array of applications.

Subsequently, Multi-Objective Optimization Algorithms, notably the NSGAII, have manifested distinctive superiority in the domain of multi-attribute optimization (Vadood & Haji, 2022b). While such algorithms can engender a compendium of approximate Pareto optimal solutions, symbolizing the compensations amongst objective functions (Boukouvalas et al., 2021), the selection of one or several pragmatically implementable solutions from this ensemble necessitates judicious deliberation by decision-makers, contingent upon specific circumstances and predilections (Boukouvalas et al., 2021).

In summation, it is discernible that each technique and methodology harbours its idiosyncratic application scenarios and intrinsic limitations. The selection of a particular technology in practical applications hinges upon a plethora of factors: the objective intended for optimization, the type and quantum of data available, the permissible computational complexity, and the requisite precision in problem resolution, inter alia. Future research might judiciously explore the amalgamation, complementation, or flexible utilization of these methodologies, contingent upon the distinctive characteristics of the problem.

Colour difference detection and assessment

For colour difference detection and assessment, the available intelligent techniques are: Support Vector Machine (SVM) algorithm and Least-squares SVM (LSSVM), Support Vector Regression (SVR) and Least-squares SVR (LSSVR), the Random Vector Functional-link net (RVFL), the Extreme Learning Machine (ELM) and the Kernel ELM (KELM) and the Regularisation ELM (RELM) and the Online Sequential ELM (OSELM) learning algorithms. Optimisation algorithms are: Principal Components Analysis (PCA) methods and GA, Whale Optimisation Algorithm (WOA) and Grasshopper Optimisation Algorithm (GOA), Sine and Cosine Algorithm (SCA) and Marine Predator Algorithm (MPA), Hunger Games Search (HGS), Takagi–Sugeno Fuzzy Neural Network (T–S FNN) and GM (1,1) grey theory models, Bagging strategies, Rotation Forest (RF), Differential Evolution (DE). Combinations of intelligent techniques and optimisation algorithms are: DE-GOA-KELM, SCA-MPA-RVFL, DPS-DE-RELM, GWO-HGS-RVFL algorithm, Bagging-PSO-ELM, RF-PSO-SLSSVR model. The 15 research articles related to colour difference detection and assessment published in the last decade are summarised in Table 2, sorted by algorithms. The elucidated intelligent techniques and optimisation algorithms furnish multifaceted solution strategies for the detection and assessment of chromatic disparities. Subsequent sections will delve into a more meticulous exploration, thoroughly scrutinizing the performance, merits, and constraints of an array of intelligent techniques and optimisation algorithms.

Table 2 Specific applications of intelligent techniques and optimisation algorithms in the field of colour difference detection and assessment

The SVM, a supervised learning methodology, primarily finds application in classification and regression issues (Zhang & Yang, 2014). SVM, along with its variant, LS-SVM, proffers a robust framework, addressing classification and regression dilemmas by harnessing the principle of structural risk minimization. Notably, the SVM operates on the principle of identifying a hyperplane, strategically maximizing the segregation of positive and negative instances (Zhang & Yang, 2014). Furthermore, LS-SVM, through the employment of equal constraints and quadratic error terms, avails a method that is computationally more expedient (Zhang & Yang, 2014). These attributes render SVM and its variants an optimal selection for addressing issues pertaining to colour discrepancies. Nonetheless, it warrants noting that the efficacy of these models is significantly contingent upon the selection of pertinent parameters, a process commonly entailing a complex cross-validation procedure.

Besides, methods based on neural networks, namely RVFL, ELM, and their variants, are to be contemplated. RVFL garners attention owing to its non-iterative training protocol, rapidity, and capacity to proffer predictive models (Liu & Yang, 2021). ELM and its variants, such as KELM and RELM, are distinguished by their straightforward architecture, swift learning velocities, and minimized parameter adjustments (Li et al., 2020; Zhou et al., 2021a, 2021b, 2019a, 2019b, 2019c). Specifically, the KELM amalgamates kernel methods with ELM, thereby enhancing performance (Li et al., 2020). Concurrently, the RELM and the OSELM respectively incorporate regularization parameters and online learning capabilities (Zhou et al., ). Nevertheless, a pervasive limitation resides in the stochastic determination of input weights and biases, potentially inducing performance instability, most notably within the online sequential variant, OSELM (Zhou et al., 2019a, 2019b, 2019c).

Turning to optimization algorithms, the PCA method is ubiquitously employed in feature extraction, celebrated for its capacity to diminish dimensions and isolate principal components (Zhang & Yang, 2014). The GA, simulating the process of natural selection, stands as a search optimization technique (Zhang & Yang, 2014). The WOA is underscored by its uncomplicated operation, scant parameter adjustments, and adeptness at evading local optima (Zhou et al., 2019a, 2019b, 2019c). The initial population configuration of the GOA may exert influence over convergence velocity and solution quality (Li et al., 2020). The SCA executes optimization searches via sine and cosine functions (Liu & Yang, 2021). However, the efficacy of these algorithms is profoundly contingent upon the selection of the initial population and parameter configurations, potentially catalyzing fluctuations in convergence velocity and the calibre of the final solution (Li et al., 2020; Zhou et al., 2019a, 2019b, 2019c).

The GM(1,1) grey prediction model distinctly characterizes itself by necessitating merely a minimal dataset for proficient modelling (Zhou et al., 2016). Conversely, Bagging, an ensemble learning methodology, efficaciously augments the performance of singular models (Zhou et al., 2016). The Rotating Forest (RF) is founded upon Random Forest and PCA (Zhou et al., 2019a, 2019b, 2019c).

Upon contemplation of the amalgamation of intelligent techniques and optimization algorithms, exemplified by DE-GOA-KELM and RF-PSO-SLSSVR, these methodologies comprehensively harness the merits of disparate algorithms, such as the expedited learning capability of KELM and the global search proficiency of GOA, in conjunction with the ensemble learning framework of RF and the parameter optimization prowess of PSO (Li et al., 2020; Zhou et al., 2019a, 2019b, 2019c; Zhou et al., 2019a, 2019b, 2019c; Zhou et al., 2019a, 2019b, 2019c). Such amalgamated methodologies strive to mitigate issues inherent in individual algorithms, such as parameter selection and local optima, whilst concurrently sustaining efficient learning and precise forecasting.

Discussions and comparative analyses pertaining to the performance, advantages, and limitations of these intelligent techniques and optimization algorithms elucidate that, within the pragmatic application domains of colour difference detection and evaluation, the selection and amalgamation of algorithms necessitate a foundation upon the inherent characteristics and prerequisites of specific issues and might necessitate subsequent algorithmic adjustments and optimizations. Prospective research endeavours, comprising the development and exploration of novel algorithms or algorithmic combinations and the investigation into the performance of varied algorithms upon disparate types of colour difference issues, will indeed present a direction of substantial value. Such endeavours will not only facilitate a more profound comprehension of the operational mechanisms inherent in various algorithms but also substantiate the development of more precise and efficient methodologies for colour difference detection and evaluation.

Colour recognition and colour segmentation

For colour recognition and colour segmentation, the available intelligent techniques and optimisation algorithms are the FCM algorithm, ANN, SVM, the K-means clustering algorithm and the X-means clustering algorithm, the Self-organising Map (SOM) and Node-Grown SOM (NGSOM), the Density Peaks Clustering (DPC) algorithm, the Efficient Dense Subspace Clustering (EDSC) algorithm, fuzzy region competition, and Fuzzy region-based segmentation (FRBS) approach. The 20 research articles related to colour recognition and colour segmentation published in the last decade are summarised in Table 3, sorted by algorithms. Subsequent sections will delve meticulously into the performance, advantages, and limitations of the aforementioned algorithms to facilitate a comprehensive comprehension of their practical applications and potential in colour recognition and segmentation.

Table 3 Specific applications of intelligent techniques and optimisation algorithms in the field of colour recognition and colour segmentation

In the realm of colour recognition and segmentation, clustering undoubtedly emerges as one of the pivotal steps (Pan et al., 2013). Among them, the unsupervised FCM algorithm is widely acclaimed as the most illustrious clustering method due to its capacity to iteratively update cluster centres and membership grades of data points, thereby positioning the cluster centres aptly within the dataset (Pan et al., 2013). Furthermore, FCM can be deployed to analyze the colour segmentation of printed fabrics within disparate colour spaces (Pan et al., 2013).

In contrast to FCM, the K-means clustering algorithm aspires to minimize the sum of squared distances between all data points and their respective cluster centres (Zhang et al., 2015a, 2015b, 2015c). Although both K-means and FCM algorithms exhibit particular merits, they are constricted by a predominant limitation, namely, the prerequisite knowledge of the number of clusters (Ouyang et al., 2019). Additionally, the X-means clustering algorithm, an extension of K-means, boasts a diminished time complexity, rendering it a significant temporal advantage when addressing problems of identical scale. For instance, the runtime of the FCM algorithm is at least 2.7 times that of X-means (Zhou et al., 2013).

The SOM, an unsupervised neural network technology, facilitates recognition or description within previously unencountered inputs (Ouyang et al., 2019). Despite the attainment of plausible colour segmentation outcomes utilizing a SOM with 10 × 10 nodes, the computational process typically proves to be time-intensive (Ouyang et al., 2019). To address this complication, the algorithm of Node Growing Self-Organizing Map (NGSOM) has been introduced, affording a reduction in processing time by up to eightfold without a compromise in accuracy (Ouyang et al., 2019).

Moreover, the DPC, a novel clustering algorithm underpinned by density, diverges from algorithms such as K-means by obviating the necessity to pre-determine the number of clustering centres (Zhang et al., 2019). This method furnishes a fresh perspective for data point classification and underscores the paramountcy of selecting clustering centres (Zhang et al., 2019).

Integration of SOM and EDSC for colour segmentation algorithms, through a bifurcated clustering approach—initially via SOM, followed by EDSC—augments the precision of segmentation considerably (Qian et al., 2022). The EDSC algorithm, inherently a subspace clustering method predicated on spectral clustering, deviates from traditional compressed sensing methods by emphasizing inter-data point connections, albeit necessitating manual input of clustering centre quantity (Qian et al., 2022).

In summation, diverse intelligent techniques and optimization algorithms each proffer unique advantages and limitations within the realms of colour segmentation and pattern recognition. The selection of an algorithm in practical applications hinges upon the specificity of task requirements and data characteristics. Future research endeavours to explore algorithms or methods capable of adaptively determining clustering centre quantity present a valuable trajectory.

Dye solution concentration and decolourisation

For dye solution concentration and decolourisation, the available intelligent techniques and optimisation algorithms are ANN, RSM, grey relational analysis (GRA), Taguchi methods, ANFIS, hSADE-NN (the ANN with a hybrid DE version), PSO algorithms, SVM, Random Forest. A summary of the 32 research articles related to dye solution concentration and decolourisation published over the last decade is given in Table 4, sorted by algorithms. The ensuing discussion proffers a comparative analysis of the efficacies of various intelligent techniques and optimization algorithms, coupled with a discourse on their respective merits and demerits.

Table 4 Specific applications of intelligent techniques and optimisation algorithms in the field of dye solution concentration and decolourisation

Primarily, ANNs furnish a potent methodology for investigating optimal conditions pertinent to the decolourization process of textile dye solutions (Khataee et al., 2013; Rosaa et al., 2013). The ANNs exhibit substantial advantages, notably their capacity to obviate the necessity for a mathematical description of the implicated process phenomena, to curtail the temporal requirements for model development, and to predict under a paucity of empirical data (Khataee et al., 2013; Rosaa et al., 2013). Subsequently, ANFIS and neuro-evolutionary methods emerged as two methodologies rooted in ANNs (Vasseghian & Dragoi, 2018). ANFIS amalgamates ANN and Fuzzy Inference System (FIS), wherein the output of each rule constitutes a linear combination of input variables, and the ultimate output manifests as the weighted mean of the output of each rule (Vasseghian & Dragoi, 2018). The neuro-evolutionary methods, on the other hand, facilitate topology and parameter optimization by conjoining ANN and Differential Evolution (DE) (Vasseghian & Dragoi, 2018). DE possesses the capability to ascertain the optimal topology and parameter values of ANN (Vasseghian & Dragoi, 2018). Moreover, hSADE-NN, an ANN enveloping a hybrid DE version, epitomizes features including a rectified mutation mechanism, an adaptive program (intended for the automatic detection of optimal control parameters), and a local search program (Vasseghian & Dragoi, 2018). Additionally, the PSO algorithm, a paradigm of search and optimization, postulates that each particle symbolizes a potential solution within the search space (Aryafar et al., 2019). Through the perpetual update of the position of each particle, predicated upon the experience and comportment of the particle and its neighbouring entities, PSO demonstrates capability in conducting searches within the problem space (Aryafar et al., 2019). The integration of PSO and neural networks empowers the efficacious determination of optimal values for weights and biases, thereby amplifying the performance of ANN (Aryafar et al., 2019).

Furthermore, GRA represents a methodology for quantifying correlations among variables, distinguished by its capacity to handle incomplete, ambiguous, and imprecise information (Zeydan, 2014). By transforming complex relationships among multiple performance characteristics into grey relational grades, GRA furnishes a simplified assessment and solution (Zeydan, 2014). Nonetheless, a potential limitation of this method may reside in its reliance upon the quality and structure of the data. The Taguchi method minimizes variations in quality characteristics by utilizing the signal-to-noise ratio (S/N) and embodies a method of experimental design applicable to the optimization of products and processes (Hossain et al., 2016a, 2016b).

Upon holistically evaluating the advantages, and limitations of various algorithms and techniques, it becomes evident that a singular method struggles to exhibit optimal performance in all circumstances. Consequently, the selection or design of algorithms in practical applications necessitates a comprehensive consideration of the characteristics and demands of the problem. The amalgamation of diverse algorithms or techniques facilitates an enhanced resolution of issues and challenges encountered in the decolourization process of textile dye solutions.

Research framework in the domain of textile colour management

In the exploration of the aforementioned domain of textile colour management, a research procedural framework was summarised, which is dedicated to shepherding researchers from the phase of problem definition through to the stage of research findings reporting (Fig. 1).

Fig. 1

General framework for applying intelligent techniques and optimization algorithms in textile colour management

The framework initially underscores the paramountcy of problem definition and objective delineation, mandating that researchers elucidate their core issues and research objectives at the inception of the project. The subsequent data collection and preprocessing phase emphasizes the acquisition and organization of data conducive to model development. Subsequently, the model selection and design phase necessitates that researchers, contingent upon the nature of the problem and characteristics of the data, select and devise suitable models. In the model training, validation, and testing phases, models are to be trained upon the data, followed by an evaluation of their performance and generalization capabilities through validation and test sets. The results analysis and optimization phase demands a thorough analysis of model outputs and, where requisite, model optimization. Once the model has undergone validation and is deemed reliable, it transitions into the actual application and evaluation phase, wherein it is deployed in a practical working environment and undergoes continuous performance assessment. Ultimately, the results reporting phase encompasses the organization and summarization of the entire research process and outcomes.

This universal procedural framework permits subdivision and applicability to four distinct sub-domains within textile colour management, each of which engages a multiplicity of intelligent techniques and optimization algorithms. For instance, in the realm of colour matching and prediction, a myriad of algorithms, inclusive but not limited to ANN, RNN, and FL, have been deployed to notable effect. In additional sectors, such as colour difference detection and evaluation, an assortment of algorithms, such as SVM, LS-SVM, and SVR, have been employed. The framework furnishes a lucid trajectory for research, enabling a direct comparison and subsequent enhancement of various intelligent techniques and optimization algorithms within each sub-domain, whilst concurrently availing opportunities for interdisciplinary research amongst the disparate sub-domains.

Performance evaluation of intelligent techniques and optimization algorithms

For the performance evaluation of intelligent techniques and optimization algorithms, researchers typically employ various metrics, including but not limited to MAE, MAPE, RMSE, and R² (Hossain et al., 2017; Li et al., 2022). The MAE primarily reflects the overall accuracy of the model and correlates closely with the consistency between model predictions and actual measurements. A model demonstrates high precision when the MAE approaches zero, indicating high consistency between predictions and actual measurements. Conversely, a larger MAE signals a requisite enhancement of model accuracy (Hossain et al., 2017). Distinctively, the Root Mean Square Error (RMSE) focuses on the comprehensive accuracy of the model. Unlike MAE, RMSE prioritizes situations with larger predictive errors, thereby capturing anomalies in predictions with heightened sensitivity (Hossain et al., 2017). Furthermore, the Mean Absolute Percentage Error (MAPE) assesses the model from a relative performance perspective, furnishing a measure of the relative size of the predictive error (Li et al., 2022). The Coefficient of Determination (R²) constitutes another pivotal metric, primarily reflecting the degree to which the model fits the data. An R² value approaching 1 denotes a high correlation between model predictions and actual observations, while a lower R² may signify suboptimal data fitting by the model (Hossain et al., 2017).

In the integration of textile colour arrangement with intelligent technology and optimization algorithms, the evaluation of model performance should not be confined to a single metric. Although foundational metrics such as MAE and RMSE can provide an initial overview of performance, ensuring the practicality of the model in actual industrial applications necessitates considering additional critical factors, such as colour difference, computational efficiency, and stability.

The pertinence of colour discrepancy attains a pivotal stance within the domain of textile colour management. The CMC formula, ubiquitously embraced, stipulates a colour variance of 1.0 unit as the permissible zenith of tolerance (Farooq et al., 2021). Progressing further, the CIE2000 formula, contrived upon a novel dataset, not only inaugurates new evaluative components but also meticulously scrutinizes the intricate relationship intertwining chroma and hue, thereby rendering a more precise reflection of tangible visual perceptions (Sharma et al., 2005).

Moreover, amidst the juxtaposition of disparate algorithms or models, cognizance of the equilibrium between computational duration and model precision becomes imperative. An algorithm, in its quest for augmented computational efficacy, might relinquish a modicum of accuracy. In contrast, a model of elevated precision might necessitate an elongated computational timespan (Chaouch et al., 2022). Stability emerges as an indispensable element. To quantify this attribute, prevalent practices involve observing the output consistency of the model amidst diverse datasets or under assorted input perturbations. Should the algorithm yield highly coherent outputs under such conditions, it could be deemed to exhibit commendable stability (Zhang & Zhou, 2022).

In summation, an integrated evaluation system, amalgamating various metrics, indubitably facilitates a more encompassing and profound assessment of a model’s overall performance, thereby furnishing researchers with lucid and comprehensive insights into its capabilities.

Results and Discussion

This section provides a descriptive statistical analysis and discussion of the 101 research articles that have been reviewed. The descriptive statistics and discussion encompass the following four sections: (1) distribution of research articles by journal; (2) distribution of research articles by year of publication and analysis of paper citations; (3) trend analysis of intelligent techniques and optimisation algorithms used in the field of textile colour management; and (4) comparison of performance of some intelligent techniques and optimisation algorithms applied in the field of textile colour management.

Distribution of research articles by journal

Figure 2 depicts the distribution of research articles among the ten domain journals. Among the 101 research articles chosen, 42 of them pertain to Materials Science Textiles (MST), accounting for 41.6% of the total. Additionally, 24 research articles belong to Chemistry Applied (CA) (23.8%) and 20 research articles belong to Engineering Chemical (EC) (19.8%). The data indicate that the selected research articles are primarily concentrated in the domains of MST, CA and EC, which are journals in the fields of textiles, chemistry and engineering, respectively.

Fig. 2

Distribution of 101 research articles selected from the Web of Science Core Collection

Distribution of research articles by year of publication and analysis of paper citations

Figure 3 illustrates the distribution of research articles over the years, demonstrating that the number of published articles (on subjects of interest) tends to fluctuate over the years. Notably, the total number of research articles published in the first 6 years (2013–2018) is 49, yielding an average of 8 research articles per year. In contrast, the total number of research articles published in the last 4 years is 52, with an average of 13 research articles published per year. This suggests the gradual proliferation and development of intelligent techniques and optimization algorithms in the realm of textile colour management. Additionally, Fig. 2 depicts the variations in the number of citations per year, indicating a transition from a relatively steady increase to a relatively rapid rise in citations per year as time progresses. The sum total of citations for the 101 selected research articles was 1126, averaging 11.15 citations per research article.

Fig. 3

Distribution of the 101 research articles by year of publication and change in the total number of citations per year of the publications

Trend analysis of intelligent techniques and optimisation algorithms used in the field of textile colour management

This section begins by describing specific intelligent techniques and optimization algorithms applied in each of the four areas: (1) colour matching and prediction; (2) colour difference detection and assessment; (3) colour recognition and segmentation; and (4) dye solution concentration and decolourization. It outlines an overview of the distribution of intelligent techniques and optimization algorithms used in the field of textile colour management during the last decade.

Figure 4 displays the main intelligent techniques used in colour matching and prediction research. ANN, FL and ANFIS have been the most commonly employed intelligent techniques, representing 64.7% and 14.7% of the 34 research articles on colour matching and prediction, respectively. The less frequently used intelligent techniques include SVM, LSSVM and RNN. The primary optimization algorithms applied in colour matching and prediction research were GA, PSO, PSO-FIMIN and ACO, accounting for 14.7%, 11.8%, and 8.8% of the 34 research articles on colour matching and prediction, respectively. Other optimization algorithms, such as MOEA, NSGAII, RSM, GWO and Taguchi's method, have been employed less frequently.

Fig. 4

Intelligent techniques and optimisation algorithms applied to colour matching and colour prediction in the last decade

Figure 5 displays the predominant intelligent techniques used for colour difference detection and assessment: (1) ELM (e.g., ELM, OSLEM, KELM, RELM); (2) SVM (e.g., SVM, LSSVM, LSSVR, SLSSVR, FWSVR); and (3) ANN, representing 40.0%, 40.0% and 33.3% respectively of the 15 research articles on colour difference detection and evaluation. RVFL, K-means algorithms, and T–S FNN have been used less frequently as intelligent techniques. The primary optimization algorithms applied for colour difference detection and assessment are DE, PSO, and RF, which represent 26.7%, 20.0%, and 13.3% respectively of the 15 research articles on colour difference detection and evaluation. respectively. The less frequently used optimization algorithms are Bagging, DPS, HGS, GA, GM(1,1), and GOA.

Fig. 5

Intelligent techniques and optimisation algorithms applied to colour difference detection and assessment in the last decade

Figure 6 demonstrates that the principal intelligent techniques used for colour recognition and segmentation are clustering algorithms (e.g. K-means, FCM, X-means), SOM, NGSOM, and DPC, representing 50.0%, 15.0%, and 10.0% respectively of the 20 research articles on colour recognition and segmentation. ANN, EDSC, and SVM have been used less commonly as intelligent techniques. The main optimization algorithm applied for colour recognition and segmentation is PCA.

Fig. 6

Intelligent techniques and optimisation algorithms applied to colour recognition and colour segmentation in the last decade

Figure 7 shows that the primary intelligent techniques employed in concentration and decolourization of dye solutions are ANN and ANFIS, accounting for 90.6% and 12.5% respectively, of the 32 research articles on the topic. SVM, on the other hand, has been employed less frequently. Regarding optimization algorithms, RSM and GA have been the primary techniques used in dye solution concentration and decolourization, while DE, GRA, PSO, RF, and Taguchi methods are less commonly applied.

Fig. 7

Intelligent techniques and optimisation algorithms applied to dye solution concentration and decolourisation in the last decade

In general, primary intelligent techniques implemented in textile colour management are (1) ANN; (2) SVM, including SVM, LSSVM, LSSVR, SLSSVR and FWSVR; (3) FL and ANFIS; (4) clustering algorithms, such as K-means, FCM, and X-means; and (5) ELM, such as ELM, OSELM, KELM and RELM. The less commonly used intelligent techniques include SOM, NGSOM, DPC, RVFL, EDSC, RNN and T–S FNN. In addition, the most frequently applied optimization algorithms are RSM, GA, PSO and DE in the selected 101 research articles. Other optimization algorithms, such as ACO, RF, GM (1,1), MOEA, NSGAII, PCA and Taguchi methods, have been used less frequently.

Performance comparison of some intelligent techniques and optimisation algorithms in the field of textile colour management

This section provides a comprehensive overview of the comparison of performances of various intelligent techniques and optimisation algorithms employed in the field of textile colour management. The comparison is based on their efficacy in tasks such as (1) colour matching and prediction; (2) colour difference detection and assessment; (3) colour recognition and segmentation; and (4) dye solution concentration and decolourisation.

Regarding colour matching and prediction, the ANN model exhibits superior predictive capabilities compared to the RSM model, as evidenced by Hasanzadeh et al. (2013), with a relative error of 1.7% for the ANN model, which is lower than the 1.8% relative error of the RSM model. Similarly, Hossain et al. (2017) and Haji and Vadood (2021) demonstrated that the ANN model provides higher accuracy in prediction than the FL model, with the ANN model achieving a higher R² (Hossain et al., 2017) and smaller RSE, MAE, and MAPE (Haji & Vadood, 2021; Hossain et al., 2017). Conversely, in Haji and Payvandy (2020), ANFIS yielded higher prediction accuracy than the ANN model, with the correlation coefficient of ANFIS being superior to that of the ANN model. Furthermore, in Vadood and Haji (2022a), using the PSO-FMIN algorithm to weight the ANN resulted in better prediction results than the BPNN, as indicated by a smaller MSE. Regarding optimisation algorithms, Chaouch et al. (2022) demonstrated that ACO can provide better predictions and higher computational efficiency than GA, as reflected by the lower mean MSE and shorter computational time required for ACO. Overall, it is important to note that different intelligent techniques may provide different levels of prediction accuracy, depending on various factors such as modulation methods, application contexts, and data structures.

Zhang and Yang (2014) used a GA-optimised SVM approach to construct an evaluation model that offers higher prediction accuracy and lower relative error than the conventional Naive Bayesian algorithm. This model provides a more reliable assessment of the quality of dyed fabrics. Similarly, Zhou et al., (2019a, 2019b, 2019c) revealed that DE-OSELM can provide superior prediction accuracy compared to the traditional textile colour correction models based on SVR and ELM algorithms. The models were also observed to exhibit better generalisation and robustness. Additionally, Zhou et al., (2019a, 2019b, 2019c) demonstrated that the DE-WOA-ELM model can yield a higher average classification accuracy than other methods, such as ELM, SVM, BPNN, and KELM, with an improvement ranging from 0.5 to 12.1%. In the study conducted by Zhou et al., (2019a, 2019b, 2019c), the RF-PSO-SLSSVR model was found to produce lower RMSE values compared to the traditional SVR and ELM algorithms. Li et al. (2020) also demonstrated that the DE-GOA-KELM method provides superior classification accuracy compared to the traditional KELM model, with an improvement of approximately 8%. Lastly, Zhou et al., (2021a, 2021b) showed that the DPS-DE-RELM model could achieve higher classification accuracy, faster convergence, and greater robustness than the conventional ELM model. From the aforementioned studies, it is evident that by combining intelligent techniques and optimisation algorithms, the prediction accuracy and robustness of traditional methods or single intelligent techniques can be enhanced.

Regarding colour recognition and segmentation, a number of studies have investigated the effectiveness of various clustering algorithms. Ouyang et al. (2019) found that the NGSOM clustering algorithm outperforms traditional SOM and FCM algorithms in terms of peak signal-to-noise ratio and time efficiency. Zhang et al. (2019) reported that the use of the DPC-SOM algorithm leads to superior predictions compared to traditional SOM, DPC, K-means, and FCM, and requires less time for execution. Additionally, Zhang et al. (2020) demonstrated that an improved version of the K-means algorithm yields higher execution efficiency than the standard K-means, FCM, and DPC algorithms. Finally, in a recent study, Das and Wahi (2022) showed that deep learning convolutional networks trained using SVM achieve higher classification accuracy than traditional BPNN. These results highlight the potential benefits of improving existing intelligent techniques or combining them with optimization algorithms to achieve better prediction accuracy or higher execution efficiency.

According to Mahmoodi et al. (2017), both RSM and GA displayed similar prediction accuracies in terms of dye solution concentration and decolourisation. Meanwhile, Vasseghian and Dragoi (2018) found that hSADE-NN demonstrated better prediction accuracy than ANFIS, resulting in a 1.5% increase in accuracy. In contrast, Nateri et al. (2019) reported that ANFIS outperforms GA, as indicated by lower MAE and MSE values and higher correlation coefficients obtained using the ANFIS model. Similarly, Baştürk and Alver (2019) found that optimising ANN using GA leads to lower prediction errors compared to the RSM model. Saleh et al. (2021) also reported higher correlation coefficients when using the ANN model compared to the RSM model. Meanwhile, Nateri et al. (2021) found that the ANFIS model produces fewer prediction errors than the ANN model. Zaharia et al. (2021) demonstrated that higher regression correlations could be achieved using the SVM model compared to the RF model. Lastly, Nnaji et al. (2022) found that, compared to ANN and RSM, the ANFIS model yields superior results with higher regression coefficients and smaller MSE. Overall, the papers cited above have indicated that the reference order for prediction accuracy ranges from high to low as hSADE-NN > ANFIS > ANN > RSM, ANFIS > GA, GA-ANN > RSM, and SVM > RF. However, it is important to note that different modulation methods and application contexts may produce different prediction accuracy results.

Future developments

Several future research directions are suggested in light of the descriptive statistics of the selected articles. Firstly, the number of research articles published on intelligent techniques and optimisation algorithms in the field of textile colour management over the last 10 years has fluctuated, with an upward trend each year, which indicates that the application of intelligent techniques and optimisation algorithms in the field of textile colour management is gradually becoming more popular and is growing fast. Thus, researchers may delve into these areas further.

Besides, when analysing the different intelligent techniques and optimisation algorithms used in the field of textile colour management, the most commonly used intelligent techniques are ANN, FL & ANFIS, clustering algorithms (e.g., K-means, FCM, X-means) and ELM (e.g. ELM, OSLEM, KELM, RELM), while the main optimisation algorithms used in the field of textile colour management are RSM, GA, PSO, DE. Furthermore, combining different intelligent techniques and optimisation algorithms or improving existing ones can provide higher prediction accuracy and faster execution efficiency than single or traditional intelligent techniques and optimisation algorithms. Therefore, future innovations may be directed at improving the existing intelligent techniques and experimenting with combinations of different intelligent techniques and optimisation algorithms to suit applications in the field of textile colour management, thus facilitating further enhancements in prediction accuracy and execution efficiency.

Furthermore, it is noteworthy that the application of intelligent techniques and optimization algorithms is not evenly distributed across the four sub-domains of the textile colour management field. For example, the traditional SVM is less frequently employed as an intelligent technique for dye solution concentration and decolourisation, whereas SVM is widely employed as an intelligent technique for colour difference detection and assessment. Furthermore, SVM has been further enhanced by LSSVM, LSSVR, SLSSVR, and FWSVR, which have the potential to provide superior prediction accuracy and faster execution efficiency. Thus, a promising future direction is to leverage the intelligent techniques and optimization algorithms presented in the four sub-domain papers to optimize the use of existing techniques and algorithms, thereby achieving improved prediction and faster convergence.

An additional innovation in the field of textile colour management could involve adapting a pre-existing algorithm via transfer learning techniques (Torrey & Shavlik, 2010). By doing so, this algorithm could be applied to other sub-domains within the field, leading to more advanced and accurate results. For instance, while ANNs are commonly used for colour matching and prediction, RNNs have shown the potential for providing more precise predictions. Thus, implementing RNNs in sub-domains where ANNs are typically used, such as colour difference detection and assessment, dye solution concentration, and decolourisation, may enhance prediction accuracy and overall efficiency.

Limitations

The present review has had some limitations concerning the selected papers. Firstly, it comprises a compilation of research articles retrieved from the Web of Science using certain keywords, focusing on the application of intelligent techniques and optimization algorithms in the field of textile colour management. This approach implies that there may be some omissions in the literature reviewed, given that the collected research articles are solely sourced from journals available in the Web of Science database, which is limited in terms of journal availability and may not cover all published research articles. Furthermore, the selection of keywords and the subsequent screening of articles performed by the authors are subjective and reliant on their own interpretation, which may also lead to some omissions. Additionally, the articles under review span from 2013 to 2022, encompassing a period of approximately 10 years, which may not provide a comprehensive outlook on the application of intelligent techniques and optimization algorithms in the field of textile colour management.

Despite the significance of intelligent techniques and optimization algorithms, their applications are limited by constraints imposed on the collection and description of research papers. For instance, due to spatial constraints, some techniques and algorithms are briefly discussed to impart a contextual understanding of their applications without providing exhaustive details. In evaluating the performance of these techniques and algorithms, results are solely derived from selected papers and are only intended for reference purposes. Predictive accuracy outcomes may differ depending on various factors, such as modulation techniques, application contexts, and data structures, necessitating further research by implementing these intelligent techniques and optimization algorithms.

Conclusions

Analysing 101 research articles from 2013 to 2022, this study comprehensively reviews the application of intelligent techniques and optimisation algorithms in textile colour management, encompassing (1) colour matching and prediction, (2) colour difference detection and assessment, (3) colour recognition and segmentation, and (4) dye solution concentration and decolourisation. The study identifies ANN, SVM and its enhancements, FL, ANFIS, clustering algorithms, and ELM, as well as optimization algorithms like RSM, GA, PSO, and DE, as widely used in this field. It finds that ANFIS often surpasses ANN and FL, ACO outperforms GA, hSADE-NN exceeds ANFIS, ANN, and RSM, and SVM is superior to RF, though results may vary based on the methods and data structures used.

Upon comprehensive analysis, the paper summarizes key research trends in intelligent technique and optimisation algorithm application in textile colour management and suggests future research could enhance existing methods, explore novel technique combinations, or leverage transfer learning to heighten prediction accuracy and optimisation efficiency.