Application of Lightweight Object Detection Network in Cucumber Leaf Detection

Abstract

The detection of cucumber leaf diseases is a pivotal link for preventing cucumber yield reduction. Manual detection lacks real-time performance and wastes a lot of human resources. Since the public dataset lacks images of cucumber leaf disease in real scenes, we collected and established a small sample dataset of cucumber leaf disease in real scenarios, and proposed a detection method of cucumber leaf disease based on YOLO v4. In order to reduce the size of the model, MobileNet v3 is used instead of CSPDarknet53 as the backbone network. By improving the residual block, introducing depthwise separable convolution, and channel pruning, the detection effect is greatly improved while reducing the network parameters and computational complexity. For small sample datasets, transfer learning is used to improve the generalization ability of the model. The model is evaluated by metrics such as FPS (frames per second), model weight size, and MAP (average of mean precision). The experimental results show that the recognition accuracy and speed of the improved model are significantly improved. MAP (Mean Precision Mean) can reach 97.21%. FPS (frames per second) can reach 40.58 frames per second. The model occupies only 25.9 MB of memory, outperforming Faster-RCNN and YOLO v4.