Convolution Neural Network Application in the Simultaneous Detection of Gender and Variety of Silkworm (Bombyx mori) Cocoons

Abstract Convolution neural network (CNN) has been widely studied in the field of computer vision with good results, and a lot of research is still needed to show its application effect in the near-infrared spectroscopy with 1-D input data and small sample set. In this paper, the near-infrared diffuse transmission spectra of 2386 silkworm (Bombyx mori) cocoons from the 4th day of cocooning to the 13th day before turning into moths were collected. After the mean center preprocessing, CNN, SVM and LDA were used to model and analyze. Experimental results showed that, with the convolution kernel width 100, the convolution kernel num 16, and the stride 50, the good generalization effect of CNN model can be achieved when the simple structure of 6 layers without pooling layer was adopted. For each day, the gender prediction accuracy of cocoons is more than 90% and the highest accuracy is 100% on the 6th day with optimal model CNN. For each day, the variety prediction accuracy of cocoons is more than 98% and the highest accuracy is 100% with optimal model LDA. There is no significant correlation between pupa age and the gender and variety detection accuracy of silkworm cocoons. This can help broaden the pupa gender identification time in the actual production of silkworm egg farms. In the application of simultaneous identification of gender and variety, CNN model has the highest accuracy of 94%, LDA model has the medium accuracy of 92.5%, and SVM model has the lowest accuracy of 89.5%.