An Instance Segmentation Method for Anthracnose Based on Swin Transformer and Path Aggregation

Abstract

Anthracnose is a common disease that affects crops, and rapid diagnosis using computer vision technology can reduce economic losses. The convolutional neural network is still the mainstream crop disease identification and disease spot segmentation method. However, the method based on the convolutional neural network is not suitable for the identification and classification tasks with minor differences between crop lesions. In view of the instance characteristics that anthracnose lesions are mostly large in size, different in size, and regular in shape, the Path Aggregation Swin Transformer Network (PAST-Net), is proposed to achieve lesion segmentation and species identification of anthracnose simultaneously. Firstly, the Swin Transformer is used as the backbone to extract features of input images. Secondly, the extracted lesion features are sequentially sent to the top-down feature pyramid network and the bottom-up augmentation path to retain the shallow network features to the greatest extent and improve the extraction ability of large-sized lesions. Next, the same proposal features from all levels are integrated by using adaptive feature pooling. Finally, the box branch performs classification and bounding box regression, while the mask branch performs lesion segmentation. The experiments show that PAST-Net improves the performance of both object detection and instance segmentation on the collected anthracnose dataset, with a recognition accuracy of 73.70% and a segmentation accuracy of 75.35%, which are 5.86% and 3.57% higher than the baseline, respectively.