Lightweight silkworm recognition based on Multi-scale feature fusion

•Application of deep learning in intensive animal husbandry and silkworm breeding.•Different from the pruning method, a lightweight method with more stable performance.•Influence of different scale feature fusion methods on model performance.•The influence of Kmeans method to reconstruct anchor boxes on dense object detection. Because the YOLOv4 model is unsuitable for the mobile and embedded terminals, YOLOv4′s lightweight MobileNetv3-YOLOv4 network significantly decreases the detection accuracy of dense silkworm targets, and the accuracy loss is too significant. A lightweight YOLOv4 detection algorithm (KM-YOLOv4) improved by multi-scale feature fusion is proposed for the target detection of dense silkworms. The Kmeans algorithm reconstructs anchor boxes suitable for different objects to enhance detection accuracy. By adding multi-scale feature fusion, the improved deep learning separable convolution MobileNetV3 lightweight backbone network replaces the YOLOv4 backbone network, reducing the computational load and model scale of the backbone network and making up for the light part of the depthwise separable convolution Accuracy loss, which improves the detection accuracy of lightweight models. The experimental results with the dense silkworm formation dataset show that the KM-YOLOv4 algorithm significantly reduces the model size by about 74% compared with the YOLOv4 algorithm and improves the detection accuracy by 1.82% with the unimproved MobileNetv3-YOLOv4 algorithm. The model can be better applied to mobile and embedded.