Research On Small Object Detection Algorithm Based On Multi-Scale Super Resolution

Small object detection is a challenging research direction in the field of object detection.Small objects are difficult to be detected correctly because of their low resolution and lack of distinguishing details,and are easy to be confused with background.In practical application scenarios,small objects are the main reason that restricts the performance of object detection algorithm,and there is great potential and space for improvement.In this dissertation,a novel super-resolution algorithm MSSR is designed for small object detection,and it is applied to improve the detection effect.Through research and experiments,it is found that the detection algorithm has an effective detection range and an optimal detection range on specific dataset.The detection range of FaceBoxes on WI)ER FACE dataset is objects larger than 32 pixels,and the best detection range is objects with sizes of about 64 pixels.For small objects beyond the detection range,there exist many scales,so it is a multi-scale problem to enlarge them to the optimal detection range.Therefore,we design a multi-scale super-resolution algorithm MSSR,which supports super-resolution reconstruction of small objects with different scales and makes them fall into the detection range.MSSR algorithm is applied to object detection using GAN.Based on the detection results of FaceBoxes,MSSR acts as a generator to generate objects with optimal size from small objects that are less than 32 pixels by super-resolution.Then,the reconstructed objects enter the discriminator for further detection.In fact,the algorithm is a two-stage extension of one-stage detection algorithm FaceBoxes.The experimental results show that the small object detection algorithm based on MSSR and GAN can improve the average precision of FaceBoxes in multiply tasks of WIDER FACE.Especially,it improves the average precision in hard task from 0.598 to 0.772,which is a significant improvement for small object detection.