Deep Learning Approaches for Multi Class Leather Texture Defect Classifcation

Leather is still a preferred item high end consumer item. It is loved for its durability and luxurious texture. However, to maintain a perfect quality standards inspection are needed for detect any imperfections lurking on the leather surface. But traditional inspection approach is very time consuming and skilled worker are required. But a deep learning approach can be an innovative approach to automating the classification of defect. Our methodology combines traditional techniques such as edge detection and statistical analysis with artificial neural networks. Different types of deep learning model are applied to compare the accuracy of the model. They are CNN, MobileNet, Softmax, DensNet, ResNet, VGG. DenseNet show the highest accuracy of 89.78%. Our model is trained and evaluated with a dateset consist of 3637 lather patches belonging to 6 classes with $227 \times 227$ pixel. This revolutionizes quality control process within the leather industry.