Research And Application On Small Object Detection Method Based On Deep Learning

Object detection is one of the most important tasks in the field of computer vision.It is the basis of other higher-level tasks(instance segmentation,pedestrian re recognition,etc.),and has been widely used in fields such as intelligent monitoring,driverless and medical image recognition.However,due to the characteristics of small objects(few pixels,fuzzy edges,etc.),the mainstream object detection algorithms for small objects still have the problems of high missed detection rate and low recognition rate.In order to solve these problems,this thesis studies the object detection algorithms based on key points,and improves the existing work from the following three aspects.(1)A feature fusion network which can effectively fuse high-level and low-level features is proposed.Since the very few pixels of small objects,it is difficult to restore specific location information after excessive downsampling.This thesis proposes a novel multi-scale feature fusion network to solve this problem,which contains multiple topdown and bottom-up feature fusion paths,the bottom-level features can effectively integrate the semantic information from the high-level,and effectively solve the problem of missing location information during downsampling.In addition,the fused features are further up-sampled to expand the resolution,which greatly enhances the feature expression ability for small objects.(2)A regression strategy based on the size and shape of the object is proposed.The previous Gaussian distribution generated by the Gaussian function for small objects is too small,usually has only one pixel,while the Gaussian distribution generated for large objects is too big,which will interfere with the training of small objects.Therefore,this thesis introduces the size factor to the original Gaussian function,which uses different rules to generate the Gaussian distribution for small objects and large objects.In addition,the shape factor is added to the Gaussian function,so that it can better represent the information of the extreme shape objects.Combining two factors can improve the training effect on small objects.(3)A knowledge distillation method for keypoint-based object detection algorithm is proposed.The heatmap generated by the teacher network is used as a mask to supervise the feature learning of the student network,so as to prevent the invalid background information from covering the information of small objects,and improve the learning efficiency of the student network.In addition,the outputs of teacher network as soft target and the sample labels as hard target supervise the training of student network together.Using the proposed knowledge distillation method to set multiple learning targets can speed up the convergence speed of the model,and make the small model achieve the accuracy close to the large model with less computational cost.Finally,this thesis designs and implements a pedestrian flow and vehicle flow density monitoring system.The detection module of the system adopts the small object detection algorithm proposed in this thesis,which verifies the effectiveness and application value of the algorithm in actual application scenarios.