A Flat Peach Bagged Fruits Recognition Approach Based on an Improved YOLOv8n Convolutional Neural Network

An accurate and effective peach recognition algorithm is a key part of automated picking in orchards; however, the current peach recognition algorithms are mainly targeted at bare fruit scenarios and face challenges in recognizing flat peach bagged fruits, based on which this paper proposes a model for recognizing and detecting flat peach fruits in complex orchard environments after bagging, namely, YOLOv8n-CDDSh. First, to effectively deal with the problem of the insufficient detection capability of small targets in orchard environments, the dilation-wise residual (DWR) module is introduced to enhance the model’s understanding of semantic information about small target defects. Second, in order to improve the detection ability in complex occlusion scenarios, inspired by the idea of large kernel convolution and cavity convolution in the Dilated Reparam Block (DRB) module, the C2f-DWR-DRB architecture is built to improve the detection ability in occluded target regions. Thirdly, in order to improve the sensitivity and precision of aspect ratio optimization, and to better adapt to the detection scenarios of targets with large differences in shapes, the ShapeIoU loss function is used to improve the fruit localization precision. Finally, we validate the effectiveness of the proposed method through experiments conducted on a self-constructed dataset comprising 1089 samples. The results show that the YOLOv8n-CDDSh model achieves 92.1% precision (P), 91.7% Mean Average Precision (mAP), and a model size of 5.73 MB, with improvements of +1.5 pp (Precision) and +0.5 pp (mAP) over YOLOv8n, respectively. In addition, the detection performance is excellent in actual orchard environments with different light angles, shading conditions, and shooting distances. Meanwhile, YOLOv8n-CDDSh deployed on the edge computing device achieved precision = 87.04%, mAP = 91.71%, and FPS = 37.20, and can also maintain high precision in bagged fruit recognition under extreme weather simulations such as fog and rainstorms, providing theoretical and methodological support for the automated picking of bagged peaches.