Super-resolution based on high-frequency feature guidance for small-scale pedestrian detection

Pedestrian detection is a long-term hot issue in the field of computer vision. In some complex scenes, such as small object size, occlusion, etc., the performance of existing detection algorithms degraded rapidly. Due to sufficient detail features are essential to improve the performance of pedestrian detection, we combine super-resolution and object detection algorithm. The existing super-resolution algorithm performs far from satisfactorily in pedestrian detection, and sometimes even brings side effects. Aiming at the above problems, we have improved the RealSR super-resolution algorithm, called <i>RealSR_plus</i>. first, A feature extraction module based on gradient shunting is adopted, Second, an improved high-frequency feature attention module based on multi-feature fusion is proposed, Furthermore, a gradient loss function based on variance improvement is used to guide the network towards the direction of optimizing the detection performance. The improved algorithm can cooperate with different detectors to achieve more accurate detection results. Extensive experiments on the challenging city persons and cyclist dataset demonstrate the effectiveness of the proposed method.