Yolov7-ODCA: Object Detection of Pomelo Epidermis Defects Based on Improved YOLOv7

The detection of pomelo epidermis defects plays an essential role in the mechanization and automation of pomelo. However, pomelo epidermis defect data possesses characteristics such as small targets and low contrasts, which brings great challenges to general object detectors. To address these issues, we propose an improved YOLOv7 network, namely YOLOv7-ODCA. The feature extraction process was improved based on the YOLOv7 network through optimization of the CBS module, and the ODCBS module was proposed by introducing the Omni-dimensional dynamic convolution (ODConv) and residual structure. This aims to enhance the feature extraction capability of CNN for low contrast images. Additionally, a Context Aggregation Block (CABlock) has been integrated in the detection head to address the issue of excessive irrelevant background information in images with small target defects. In the experiments, we compare the proposed YOLOv7-ODCA networks in the metrics of AP, AP50, precision, and recall. It achieves average accuracies exceeding all compared networks, with values of 49.8%, 92.6%, 87.6%, and 85.2%, respectively.