Heterogeneous image features are complementary, and feature fusion of heterogeneous images can increase position effectiveness of occluded apple targets. A YOLOfuse apple detection model based on RGB-D heterogeneous image feature fusion is proposed. Combining the CSPDarknet53-Tiny network on the basis of a YOLOv5s backbone network, a two-branch feature extraction network is formed for the extraction task of RGB-D heterogeneous images. The two-branch backbone network is fused to maximize the retention of useful features and reduce the computational effort. A coordinate attention (CA) module is embedded into the backbone network. The Soft-NMS algorithm is introduced, instead of the general NMS algorithm, to reduce the false suppression phenomenon of the algorithm on dense objects and reduce the missed position rate of obscured apples. It indicates that the YOLOfuse model has an AP value of 94.2% and a detection frame rate of 51.761 FPS. Comparing with the YOLOv5 s, m, l, and x4 versions as well as the YOLOv3, YOLOv4, YOLOv4-Tiny, and Faster RCNN on the test set, the results show that the AP value of the proposed model is 0.8, 2.4, 2.5, 2.3, and 2.2 percentage points higher than that of YOLOv5s, YOLOv3, YOLOv4, YOLOv4-Tiny, and Faster RCNN, respectively. Compared with YOLOv5m, YOLOv5l, and YOLOv5x, the speedups of 9.934FPS, 18.45FPS, and 23.159FPS are obtained in the detection frame rate, respectively, and the model are better in both of parameter’s number and model size. The YOLOfuse model can effectively fuse RGB-D heterogeneous source image features to efficiently identify apple objects in a natural orchard environment and provide technical support for the vision system of picking robots.
Loading....