Rice Ears Detection Method Based on Multi-Scale Image Recognition and Attention Mechanism

Abstract

Accurate identification of rice ears is crucial for assessing rice yield. Present research mainly relies on single-scale image data for rice ears detection and counting. However, these approaches are susceptible to misdetection and omission due to the intricate environmental conditions in fields. The combination of multi-source images can better overcome the limitations of single-scale images. In this study, based on the YOLOv5s target detection algorithm, a method for rice ears detection and counting applicable to multi-source images is proposed by integrating image data collected by cell phones and UAVs during the rice heading and maturity periods. The proposed method introduces Attention-based Intrascale Feature Interaction (AIFI) to reconstruct the backbone feature extraction network, optimizing feature expression interaction and enhancing handling of the model of advanced semantic information. Additionally, Simplify Optimal Transport Assignment (SimOTA) is employed to achieve a more refined label assignment strategy, thereby optimizing detection performance of the model and addressing difficulties in detecting multiple targets in high-density rice ears environments. Finally, Channel-wise Knowledge Distillation for Dense Prediction (CWD) is utilized to enhance the performance of the model in dense prediction tasks by transferring knowledge between different channels. The experimental results demonstrated good performance of the model on datasets comprising rice at the heading and maturity stages, achieving Precision, Recall, and mAP values of 93%, 85.3%, and 90.3%, respectively. The coefficients of determination (R 2 ) for the linear fit between test results and the actual statistical results of the model were 0.91, 0.91, 0.90, and 0.88, respectively. The proposed model performs well in the mixed dataset and can be utilized more effectively for accurate identification and counting of rice ears.