关键词: 油茶果, YOLOv5s, BSAM注意力机制, 可改变核卷积, 软性非极大值抑制

Abstract: Aiming at the current problems such as lower recognition accuracy and higher leakage rate of camellia oleifera fruit in complex environments, we propose a camellia oleifera fruit detection algorithm based on improved YOLOv5s, BAS-YOLOv5s. The algorithm utilizes the BSAM attention mechanism in the backbone network to improve the feature extraction capability of the network and to im-prove the recognition of small and occluded targets by the original network. Second, to allow the convolution kernel to be irregular and to improve the generalization of the model, alterable kernel convolution is introduced into the structure. Finally, Soft-NMS is used instead of NMS in YOLOv5s to re-filter the target frames and reduce the probability of overlapping targets being mistakenly deleted. The improved model is also cross-validated with K-fold, and the average F1 value reaches 88.6%. The validation of the photographed dataset shows that the average precision of the improved model reaches 92.3% and the recall rate reaches 84.3%, which is 2.0% and 1.9% higher than the original YOLOv5s algorithm, respectively, and the FPS reached 82.64 frames per second, which is better than other main-stream algorithms in general, and proves the effectiveness of the improved method for the recognition of camellia oleifera fruit in the complex environment. 

Key words: camellia oleifera fruit, YOLOv5s, bilevel spatial attention module, alterable kernel convolution, Soft-NMS