In this research,an improved CenterNet model(SEP-CenterNet)combining Efficient Channel At-tention(ECA)mechanism for rice appearance quality detection was proposed,aiming to solve the problem of low recognition accuracy in existing detection methods.Firstly,images of broken rice,whole rice,and yellow rice were obtained using image acquisition equipment,and data augmentation techniques,such as rotation and flipping,were applied to prevent overfitting due to small datasets.Then,GhostNet with Spatial Pyramid Pooling(SPP)was used as the backbone network of CenterNet to extract multi-level features of rice,and a Path Aggregation Network(EPA-Net)based on ECA attention mechanism was used for semantic feature fusion.Experimental results indicated that the detection accuracy of the improved CenterNet model for broken rice,yellow rice,and whole rice reached 97.02％,96.73％,and 98.14％,respectively,with anmAP improvement of 5.84％compared to the original CenterNet mod-el.The recognition accuracy was higher than that of other models,such as SSD,Faster-RCNN,Retinanet,YOLOV4,YOLOV4-tiny,and YOLOV5.Moreover,compared with the traditional rice detection methods based on extracting morphological features and Fourier coefficients,the proposed model had higher accuracy and better general-ization performance.
关键词
大米外观品质/目标检测/空间金字塔池化/路径聚合网络/注意力机制
Key words
appearance quality of rice/target detection/spatial pyramid pooling/PANet/attention mechanism
维普
万方数据