Development of an Optimized Deep Learning Technique for Fabric Defect Classification Using Osprey Optimization Algorithm

Abstract

Fabric defect classification system lowers the expenses related to low-quality products by improving productivity and product quality. There is quite a number of existing fabric defects classification techniques but most of them are not suitable for real-time classification in complex environment due to insufficient deep feature extraction. Hence, this research developed an enhanced deep learning technique using Osprey Optimization Algorithm (OOA) for fabric defect classification because of the strong eye sight of the osprey bird the algorithm is model to simulate. Fabric defects datasets were obtained from kaggle.com. Three thousand, seven hundred and ninety three (3793) datasets were obtained: two thousand six hundred and fifty five (70%) and one thousand one hundred and thirty eight (30%) were used for training and testing, respectively. The acquired datasets were pre-processed in MATLAB using k-means clustering for segmentation of region of interest, Gaussians filter for noise removal and Contrast Limited Adaptive Histogram Equalization (CLAHE) for image contrast enhancement. Densenet Cross Stage Partial Network Darknet53 (DCSPDarkNet53) was Optimized with Osprey Optimization Algorithm. The resulting OspreyDCSPDarkNet53 was used to extract features from the pre-processed datasets. The developed OspreyDCSPDarknet53 was implemented with MATLAB (R2022) and its performance was evaluated using False Positive Rate (FPR), Specificity, Sensitivity, and Accuracy. The developed method was compared with DCSPDarknet53 and YOLOv4 methods. The FPR, Specificity, Sensitivity, and Accuracy for OspreyDCSPDarknet53 were 3.30%, 96.70%, 96.09%, and 96.40%, respectively while the FPR, Specificity, Sensitivity, and Accuracy for DCSPDarknet53 were 4.86%, 95.14% 94.48%, and 94.82%, respectively and the FPR, Specificity, Sensitivity, and Accuracy for YOLOv4 were 7.47%, 92.53%, 91.81%, and 92.18%, respectively. The developed OspreyDCSPDarknet53 performed better than other techniques in terms of all the evaluated metrics. Hence, it can find its application in real time classification of fabric defects.