Research On The Small Object Detection Algorithm Based On Deep Learning

In recent years,deep learning technology has developed rapidly,which has promoted the target detection algorithm.Compared with the traditional algorithms,the deep learning-based object detection algorithm is more robust.As an important branch of object detection,small object detection has important research significance in UAV scene analysis,pedestrian detection,unmanned driving,and other applications.However,the small proportion of small target pixels makes the detection effect far lower than large/medium scale targets.The detection difficulties are:(1)small targets are susceptible to the deviation of prediction frame and low positioning accuracy;(2)small targets are poor anti-interference and susceptible to target occlusion;(3)small target regression offset and the number of samples in the data set is small,resulting in low loss contribution.In this paper,the aspect ratio of the target annotation frame and image width is less than 0.1as the benchmark of small target determination,and the above problems existing in the medium and small target detection are carried out in visual tasks.The main research contents are as follows:(1)In view of the low positioning accuracy of the deep learning object detection algorithm,this paper proposes a small object detection algorithm based on the adaptive class attribute ambiguity perception(ACBA).First,ACBA is defined as the absolute value of the Io U and the median value of each sample,so as to perceive the attribute ambiguity of the positive and negative samples of the target class.Secondly,based on cross-entropy loss and ACBA,it is constructed to suppress the low confidence information through class attribute ambiguity and alleviate the inundation problem of confidence information.Finally,the constraint factor is designed to adapt the dependence of the single regression network on confidence information,and the penalty factor is designed to balance the proportion of attribute ambiguity in the loss.In the VOC2007 dataset,ACBA Loss improved the YOLO V4 average accuracy(m AP)and small target average accuracy(AP)by 1.77% and 10.5%,respectively.(2)In response to the problem of poor anti-interference ability of small targets in deep learning object detection algorithms and the significant impact of target occlusion,this paper proposes a small target detection algorithm based on anchor box multi-target matching degree deviation to improve the robustness of the model in detecting occluded targets.First,fuzzy samples are defined as samples overlapping with multiple annotation boxes and are also the target where occlusion may occur.Secondly,the multi-target matching degree deviation(MTMD)is defined by the two maximum Io U differences of the fuzzy sample matching multi-annotation box,as the interference degree of the quantification of the prediction box matching fuzzy sample.Then,the adaptation weight is constructed based on the MTMD,and the equilibrium parameter Φ for the non-fuzzy sample,and the multi-target matching deviation loss(MTMD Loss)is obtained.Finally,MTMD-F Loss is proposed between MTMD Loss and Focal Loss to improve the equilibrium performance of the model by simultaneously focusing on fuzzy samples and difficult samples.In the Widerperson dataset with severe target occlusion,MTMD Loss and MTMD-F Loss increased the m AP of YOLO V4 by 2.04% and 2.36%,respectively.In the VOC2007 dataset,MTMD Loss and MTMD-F Loss increased the m AP of YOLO V4 by 1.62% and 2.45%,respectively,and the small target AP by 9.8% and 6%,respectively.(3)For the low contribution of small object loss,this paper proposes a small object detection algorithm based on target area normalization.First,the annotated box area was obtained by walking through the target label in each training sample picture.Second,each annotation box area was normalized by Min-max and assigned to the corresponding prediction box,reflecting the sample contribution.Then,the classification and location loss weights based on the normalized area normalized loss(TAN Loss).Finally,TAN-F Loss is proposed between TAN Loss and focus loss(Focal Loss)in the classification loss branch to improve the equilibrium performance of the model by simultaneously focusing on small target samples and difficult samples.In the VOC2007 data set,TAN Loss and TAN-F Loss increased the m AP of YOLO V4 by 1.75% and 2.73%,respectively,and the small target AP by 7.5% and 7.3%,respectively.