ECBFE-YOLOv5: Efficient Cross Branch Feature Extraction Module for Shaped object Detection

Shaped objects have complex shapes and structures that are difficult to capture and characterize compared to traditional regular shapes. In addition, since the appearance and size of shaped objects may change with the observation angle, this further increases the difficulty of target recognition. In this paper, a novel target detection algorithm ECBFE-YOLOv5 is proposed for recognizing shaped objects based on YOLOv5, which reconstructs the feature extraction module in the original YOLOv5 algorithm. The network accuracy is improved by changing the residual network and spatio-temporal feature fusion path. This can enrich the different levels of feature information to achieve the purpose of the new model’s accuracy improvement. Through comparative experiments, the map0.5 and map0.5:0.95 of our proposed ECBFE-YOLOv5 are improved by 0.9% and 1.2%, respectively. The experimental results demonstrate the feasibility of ECBFE-YOLOv5.