Model for the online frontal image selection of silkworm pupae using machine vision

The sorting of male and female silkworm pupae is an essential process of silkworm breeding, with its accuracy directly affecting the quality of hybrid silkworm eggs and silk. Gonadal characteristics serve as a reliable basis for sex identification in silkworm pupae; however, the gonads only exist on the positive side of the tail. Due to the unique geometry of silkworm pupae, online sex recognition based on machine vision requires flipping and taking many photos of the same silkworm pupae. Thus, accurately selecting the frontal image from multiple images of the same silkworm pupae in different poses is a prerequisite for subsequent sex identification. To address this challenge, we proposed SPNet-GS (Silkworm Pupae Network for Gonad Selection), a lightweight model for online selection of frontal silkworm pupae images. The model first employed a large kernel convolution to enhance the receptive field and capture the relevant information between adjacent pixels. Then the correlation between long-distance pixels under multi-scale information can be obtained by dilated convolutions. Finally, the correlation information between near and far pixels was fused to enhance feature extraction. Experimental results demonstrated that our method outperforms other models with an average accuracy of 98.41% and an average F1 score of 99.02%. The average inference time of each image was 0.03 s, which can fully meet the requirements of online selection of male and female silkworm pupae. Moreover, the gender identification accuracy rates using the selected frontal image and gonad region image reached 84.68% and 94.58%, respectively. These results were 10% and 19.90% higher than using multi-pose images for sex identification, demonstrating the effectiveness of the frontal image selection strategy. The findings of this investigation may provide a valuable reference for the machine vision-based intelligent online sorting of silkworm pupae by gender.