Method of Locating the Strike Point on Pest for Laser Control Based on Mask R-CNN

To achieve pests elimination on the cabbage leaves by laser, it is necessary to accurately identify and locate the laser strike points on the pests. This paper, takes Pieris rapae (Lepidoptera: Pieridae) as the research object, a method for laser strike point localization was proposed. Firstly, an image acquisition device of 850 nm wavelength was constructed and acquired the near-infrared image simples for the dataset. Then, the segmentation of pest targets was performed using different backbone models to enhance the Mask R-CNN. Additionally, the centerline of pest body was extracted by the improved Zhang-Suen thinning algorithm to determine the chain code. Eventually, the midpoint of the centerline which was considered in the middle area of pest abdomen, was determined as the laser strike point. As the experimental results show, the evaluation metrics average precision (AP) of all models (ResNet50, ResNet101, and Inception-v2) exceeded 90% on the near-infrared image dataset, and the AP, AP0.5, and AP0.75 reached 94.24%, 98.74%, and 96.79% with the ResNet50-based Mask R-CNN model. The maximum location error in the x-coordinate, y-coordinate, and the distance between the real and the located point, were 6, 5, and 6.32 pixels, respectively, and the root mean square error of the total location error was 3.35 pixels. The research would provide the necessary reference for developing the laser pest control robots.