A Lightweight Convolutional Neural Network for Silkworm Cocoons Fast Classification

Silk clothing is very popular all over the world, and the annual demand is very large each year. Silkworm cocoons as raw materials for silk will be selected much more to meet market demand. At present, high quality cocoons and defective cocoons are identified mainly by manual selection, the main problem of this situation is that the identification process is inefficient. It is necessary to develop automated visual equipment to assist workers in silkworm cocoons selection. In recent years, deep learning especially the convolutional neural network (CNN) has achieved high success and gradually become the main method in the area of image classification. In order to establish silkworm cocoon classification model, we made a silkworm cocoon data set and designed a convolutional neural network to classify silkworm cocoons. For traditional convolutional neural networks, the fully connection layer often contains most of the parameters of the whole model, which consumes a lot of computing resources and time in the forward propagation. In order to improve this, we use global average pooling (GAP) layer to connect the convolutional layer and classify layer to reduce some fully connection layers. And all improvements were accomplished on the basis of the structure of AlexNet. As a result, the AlexNet (GAP) is obtained. Experimental results demonstrate that the proposed architecture achieves a mean accuracy of 98.22%, which makes it possible to apply cocoons classification in industry very soon.