Abstract: Aiming at the problem that the YOLO_v4 model would be gradually  diluted due to the features of serial connection of the neck network, which affected the performance of the model, we proposed a YOLO_v4 object detection method with improved multi-scale features. The method  reconstructed the YOLO_v4 neck network structure   by introducing an intermediate layer, and then used the intermediate layer to participate in the subsequent feature fusion to realize the cross-level connection of features, and used the parameters that could be learned through the network as the balance factor between the features to perform  feature weighted fusion. The experimental results on the VOC-2007 and VOC-2012 datasets show that the proposed  method  can improve the average accuracy of the model by 1.3%, which can effectively improve the detection ability of the model for different targets.

Key words: object detection, deep learning, multi-scale feature, weighted fusion