Deep Learning based Feature Extraction for Texture Classification

Abstract

Categorizing Texture plays a central role in performing automated machine vision tasks such as defect detection and visual inspectionin industries and factories. Classifying texture is a prominent step in pattern recognition problems. Hand Crafted Texture features or Texture descriptors are found successful in identifying and classifying different textures. Deep learning based techniques are also competent enough to categorize and identify texture images. Texture classification is a computer vision task which is applied in industrial applications such as visual inspection, fabric defect detection, automatic PCB fault detection etc. Deep Network requires enormous amount of data for training and it is extremely memory intensive. In this paper, we propose methods where Convolution Neural Network (CNN) features are used for feature extraction and Support Vector machine is used as classifier for texture classification. We used cross entropy as the loss function to estimate the error during training. We investigate the efficiency of using CNN features extracted from the different pretrained models DenseNet201, ResNet50, ResNet101, Inceptionv3, AlexNet and classifying using SVM classifier. Performance is computed in gray and color texture databases such as KTH-TIPS, CURET and flower datasets. Results show good and superior accuracy of about 85%-95% with different datasets. These proposed methods have less computation time.