DC-YOLOv4 algorithm based on improved multi-scale adaptive feature fusion

According to the invention, a YOLOv4 improvement method is researched mainly aiming at a multi-scale target coexistence phenomenon of small target detection. Although the multi-scale feature fusion mode of the PAFPN increases the overall detection precision of the target, the second bottom-up feature transfer does not provide more small target features. According to the invention, an original PAFPN multi-scale feature fusion structure is adjusted, firstly, jump connection is used for transmitting same-level feature maps, then, shallow-layer feature maps are transmitted to the upper layer, small targets are conveniently detected, finally, a CARAFE up-sampling operator is introduced into the first-time top-down structure, context information inflow is increased, and therefore, the multi-scale adaptive feature fusion algorithm based on improvement is obtained. According to the invention, the algorithm model is called DC-YOLOv4, and experiments prove that the small target detection precision can be effectively im