A Small Target Detection Method Based on Feature Enhancement and Positioning Optimization

Abstract

Feature Pyramid Network (FPN) is a basic but important component in target detection system. Together with target detection algorithms, such as SSD, Faster R-CNN and YOLO series, they have achieved good detection results for large targets with high resolutions, but the performance is less effective when it comes to detect small targets that contain relatively little semantic information. And small target detection is quite common in daily life, such as face recognition at long distances, traffic sign detection in automatic driving, etc. That means it is significant to break the bottleneck of target detection and get better accuracy performance. In this article, based on the FPN, we propose an improved network structure (IMFPN) from two aspects to get a better accuracy result in small target detection task. In the first aspect, we improve the feature map pyramid structure for feature enhancement, reduce the problem of information loss during feature map fusion and get the semantic information of multiscale feature maps. In the second aspect, we concentrate on the problem of information loss in the pooling process of feature maps, we propose an improved version of the PRRoI pooling method that combines RoI Pooling and RoI Align Pooling. And we also optimize the positioning of the frame through a new IoU calculation standard. Based on these above ideas and methods, we propose a small target detection method based on feature enhancement and positioning optimization.