RSF-SSD: An Improved SSD Algorithm Based on Multi-level Feature Enhancement

SSD (Single Shot Multi Box Detector) is one of the commonly used object detection algorithms known for its fast detection speed and high accuracy. However, SSD's performance in detecting objects of different scales is suboptimal. This paper proposes the RSF-SSD network based on multi-level feature enhancement. By improving the backbone network of SSD, skip-connections and channel attention mechanism are introduced into VGG16. This operation enhances the ability of the backbone network to extract detailed features. In the feature fusion module, an improved FPN (Feature Pyramid Networks) + PAN (Path Aggregation Network) module is introduced (referred to as FAN) to achieve comprehensive fusion of deep semantic information and shallow localization information. In testing on the PASCAL VOC07 + 12 dataset, the proposed network achieves a 15.3%mAP improvement in detection accuracy compared to traditional SSD. The experiments demonstrate that the RSF-SSD model effectively enhances the capability of extracting detailed features, and incorporates more semantic information in the shallow layers, leading to improved detection performance for objects of different sizes.