Small object detection network based on YOLOv8

The small size, limited feature, and low contrast of small objects challenge the network's capacity to extract and encode object attributes, thereby diminishing the efficacy of small object detection networks. This problem is solved by utilizing a YOLOv8-based object detection network for the detection of small objects. At the beginning, the pyramid vision Transformer (PVT) backbone network is deployed to elevate spatial feature extraction efficiency and reduce model parameters; then, a multi-scale adaptive feature fusion (MAFF) architecture is crafted to amplify the representational capability of multi-scale spatial features. The experimental outcomes demonstrate an improvement in the AP, AP0.5 and AP0.7 through the proposed strategy. Secured on the AI-TOD dataset, via an 8.6%, 21.2%, and 5.6% in comparison to the baseline, the model's small object detection proficiency was notably improved.