Improving Object Detection in YOLOv8n with the C2f-f Module and Multi-Scale Fusion Reconstruction

In the field of object detection, such as the YOLO series, the backbone network has been significantly optimized, but there needs some improvement for the space feature extraction. For example, when small objects are detected, there is a high probability of missed detection. In order to address this problem, a novel object detection module is introduced, and called C2f-f module.Firstly, the receptive fields of various features have been extended during the expansion convolution process to discover the differences between detailed target features and their relationships based on parameters. Secondly, when the necessary low-level feature information is extracted, also a majority of high-level information is preserved to ensure and mitigate the continuity of information flow, also prevent the gradient from disappearing. Thus, this helps to extract residual high-level information and improve the accuracy of the related model. Finally, a reconstruction for the backbone network based on a multi-scale fusion method with the principles of BiFPN is proposed to improve the network structure of YOLOv8n. Experimental validation on the COCO dataset demonstrates the effectiveness of the proposed method. More precise detection results for small object tasks are obtained, also the universality and innovation are exhibited via the new method than the existing ones.This new module exhibits more precise detection results for small object tasks, showcasing its universality and innovation across different models.