Abstract: Aiming at the problems of dense detection targets, occlusion, missed detection caused by inconsistent morphology in various regions and weak generalization ability of the model in the wheat ear dataset, we proposed a wheat ear detection algorithm based on improved YOLOv7. Firstly, we introducd a mixed attention mechanism into the backbone feature extraction network of YOLOv7 network to strengthen the extraction of location features and alleviate the missed detection problem caused by dense detection targets. Secondly, switchable atrous convolution (SAC) which could combine different sizes was introduced into the backbone feature extraction network, and the feature information of different scales was extracted by increasing the receptive field, which could effectively improve the missed detection problem caused by occlusion. Finally, an incremental learning module example vector correction (EVC) was introduced into the feature fusion part to improve the robustness and generalization ability of the model. The experimental results show that the average target detection accuracy of the improved wheat ear recognition algorithm in the global wheat ear dataset is 2.11 percentage points  higher than that of the original YOLOv7.

Key words: wheat ear detection, mixed attention, incremental learning, atrous convolution