CMA-YOLO: A Wine Grape Detection Model Based on YOLOv5x Combining Mixed Attention Mechanism

Abstract

The progress of object detection technology can promote the automation and modernization of agricultural production. In order to improve the feasibility and efficiency of automated fruit harvesting, this study proposes an improved object detection model based on YOLOv5x. Firstly, considering the lack of training samples, we improved the data loading method by adding an input branch for grayscale processing. Combined with the mosaic data augmentation method, the diversity of training samples has been dramatically improved, enriching the training set and significantly improving training effectiveness. Secondly, we add a global self-attention mechanism to the middle layer of the backbone network to highlight essential features and suppress irrelevant features, making the distinction between targets and the background more obvious. Finally, we recombined channel and spatial attention in serial and parallel methods. We integrated them into the C3 module of the neck section of the model to obtain a new CMA-C3 module, effectively improving the model's attention to features. Compared with the baseline network, our model only adds minimal complexity, but the detection precision, F1-score, mAP@0.5, and mAP@[0.5:0.95] have improved by 1.7%, 3.2%, 8.7%, and 10.2% respectively. In addition, we have verified the effectiveness and generalization of the improved modules and schemes through multiple ablation experiments and confirmed that the proposed improved model could achieve accurate object detection of wine grapes, providing fast and accurate guidance for the automatic harvesting of grapefruits and promoting the automation and intelligence of orchard agricultural production.