CA-YOLOv5: Detection model for healthy and diseased silkworms in mixed conditions based on improved YOLOv5

The accurate identification and localization of diseased silkworms is an important task in the research of disease precision control technology and equipment development in the sericulture industry. However, the existing deep learning-based methods for this task are mainly based on image classification, which fails to provide the location information of diseased silkworms. To this end, this study proposed an object detection-based method for identifying and locating healthy and diseased silkworms. Images of mixed healthy and diseased silkworms were collected using a mobile phone, and the category and location of each silkworm were labeled using LabelImg as a labeling tool to construct an image dataset for object detection. Based on the one-step detection model YOLOv5s, the ConvNeXt-Attention-YOLOv5 (CA-YOLOv5) model was designed in which the large kernel with depth-wise separable convolution (7x7 dw-conv) of ConvNeXt was adopted to expand receptive fields and the channel attention mechanism ECANet was added to enhance the capability of feature extraction. Experiments showed that the mean average precision (mAP) values of CA-YOLOv5 for healthy and diseased silkworms reached 96.46%, which is 1.35% better than that achieved via YOLOv5s. At the same time, the overall performance of CA-YOLOv5 was significantly better than state-of-the-art one-step models, such as Single Shot MultiBox Detector (SSD), CenterNet, and EfficientDet, and even improved YOLOv5 using image attention mechanism and a lightweight backbone, like SENet-YOLOv5 and MobileNet-YOLOv5. The results of this study can provide an important basis for the accurate positioning of diseased silkworms in precision disease control technology and equipment development.