Cashmere and Wool Classification with Large Kernel Attention and Deep Learning

Abstract

Aiming at the problems of low manual detection efficiency and low automatic detection accuracy of cashmere and wool, a method is proposed to realize the image detection of cashmere and wool single fiber by using large kernel attention(LKA) mechanism and deep convolutional neural network. Based on the ConvNeXt network structure paradigm with simple structure, good scalability and high accuracy on ImageNet large datasets, the inverted residual structure improvement avoids information loss, and large kernel attention mechanism is added to make the model more accurate to pay attention to differentiated regions on the feature map space and channel, while considering the sparse amount of fiber image information, in order to avoid redundant training parameters and overfitting, the network is lightweight while maintaining the proportion of the ConvNeXt network hierarchy, and the LKA-RConvNeXt model is established. Finally, after training on 15,000 cashmere and wool datasets, the highest classification accuracy can reach 96.1%. Through ablation experiments and model comparison analysis, it is verified that the improved method used is beneficial to the accuracy of the model. The model can be used for cashmere and wool in automatic classification tasks, and contributes to backbone network for the subsequent fiber object detection task.