Research on Object Detection Model Based on Feature Network Optimization

As the object detection dataset scale is smaller than the image recognition dataset ImageNet scale, transfer learning has become a basic training method for deep learning object detection models, which pre-trains the backbone network of the object detection model on an ImageNet dataset to extract features for detection tasks. However, the classification task of detection focuses on the salient region features of an object, while the location task of detection focuses on the edge features, so there is a certain deviation between the features extracted by a pretrained backbone network and those needed by a localization task. To solve this problem, a decoupled self-attention (DSA) module is proposed for one-stage object-detection models in this paper. A DSA includes two decoupled self-attention branches, so it can extract appropriate features for different tasks. It is located between the Feature Pyramid Networks (FPN) and head networks of subtasks, and used to independently extract global features for different tasks based on FPN-fused features. Although the DSA network module is simple, it can effectively improve the performance of object detection, and can easily be embedded in many detection models. Our experiments are based on the representative one-stage detection model RetinaNet. In the Common Objects in Context (COCO) dataset, when ResNet50 and ResNet101 are used as backbone networks, the detection performances can be increased by 0.4 and 0.5% AP, respectively. When the DSA module and object confidence task are both applied in RetinaNet, the detection performances based on ResNet50 and ResNet101 can be increased by 1.0 and 1.4% AP, respectively. The experiment results show the effectiveness of the DSA module.