Learning Discriminated Features Based on Feature Pyramid Networks and Attention for Multi-scale Object Detection

As the research scene in object detection becomes increasingly complex, the extracted feature information needs to be further improved. Many multi-scale feature pyramid network methods have been proposed to improve detection accuracy. However, most of them just follow a simple chain aggregation structure, resulting in not considering the distinction between multi-scale objects. Modern cognitive research presents that human cognitive ability is not a simple image-based matching process. It has an inherent process of information decomposition and reconstruction. Inspired by this theory, a new feature pyramid network model denoted as SuFPN based on discriminative learning is proposed to solve the problem of multi-scale object detection. In SuFPN, the correlation between the underlying location information and the deep feature information is fully considered. Firstly, object features are extracted through top-down path and lateral connection. Then deformable convolution is used to extract object discriminant spatial information. Finally, the attention mechanism is introduced to generate a discriminative feature map with enhanced spatial and channel interdependence, which provides excellent location information for the feature pyramid while considering semantic information. The proposed SuFPN is validated on the PASCAL VOC and COCO datasets. The Average Precision (AP) value reaches 80.0 on the PASCAL VOC dataset, which is 1.7 points higher than the feature pyramid networks (FPN), and 39.2 on the COCO dataset, which is 1.8 points higher than the FPN. The result demonstrates that our SuFPN outperforms other advanced methods in the multi-scale detection precision.