CMRD-Net: a deep learning-based Cnaphalocrocis medinalis damage symptom rotated detection framework for in-field survey

The damage symptoms of Cnaphalocrocis medinalis (C.medinalis) is an important evaluation index for pest prevention and control. However, due to various shapes, arbitrary-oriented directions and heavy overlaps of C.medinalis damage symptoms under complex field conditions, generic object detection methods based on horizontal bounding box cannot achieve satisfactory results. To address this problem, we develop a Cnaphalocrocis medinalis damage symptom rotated detection framework called CMRD-Net. It mainly consists of a Horizontal-to-Rotated region proposal network (H2R-RPN) and a Rotated-to-Rotated region convolutional neural network (R2R-RCNN). First, the H2R-RPN is utilized to extract rotated region proposals, combined with adaptive positive sample selection that solves the hard definition of positive samples caused by oriented instances. Second, the R2R-RCNN performs feature alignment based on rotated proposals, and exploits oriented-aligned features to detect the damage symptoms. The experimental results on our constructed dataset show that our proposed method outperforms those state-of-the-art rotated object detection algorithms achieving 73.7% average precision (AP). Additionally, the results demonstrate that our method is more suitable than horizontal detection methods for in-field survey of C.medinalis.