Camouflaged cotton bollworm instance segmentation based on PVT and Mask R-CNN

Abstract

Many pests change their appearance color to seamlessly blend with the surrounding environment in agricultural ecosystems, thereby rendering themselves virtually invisible. When the pest’s color and texture resemble the background, accurately identifying and detecting it becomes challenging. In this study, we construct a new dataset focusing on the cotton bollworm and conduct in-depth optimization and improvement of the instance segmentation model based on the Pyramid Vision Transformer (PVT) and Mask R-CNN. To better capture the features of camouflaged organisms, the proposed model utilizes the PVT as a feature extraction network and Mask-RCNN for instance segmentation. We also introduce overlapping image embedding patch structure and further incorporate a feed-forward network with depthwise separable convolution. These improvements enhance the PVT’s capability to capture global and intricate features and significantly boost the accuracy of instance segmentation. Considering the computational efficiency demands in real-time agricultural applications, we introduce a linear spatial-reduction attention mechanism that effectively reduces computational complexity. The experimental results show that the model achieves the detection accuracy of 89.7% and the segmentation accuracy of 89.2% for camouflaged cotton bollworms.