Research On Small Object Detection In Complex Scenes

Object detection is an important research direction in computer vision,and also the basis for other complex visual tasks.In recent years,with the rapid development of deep learning,object detection algorithms have made a huge breakthrough.At present,object detection has been widely used in many fields,including automatic driving,intelligent monitoring,and medical image analysis.However,the existing algorithms can not deal well with small objects,especially in complex scenes.In the real scene,the size of the object is usually small,and it is usually accompanied by the dramatic changes of illumination,partial occlusion and scale change.Therefore,small object detection in complex scenes is very challenging.In this thesis,a series of innovative researches on small object detection in complex scenes are carried out.Based on the existing object detection algorithms,a variety of novel methods are proposed to improve the performance of small object detection.The main contributions of this thesis are summarized as follows:1.To solve the problem of small object feature extraction,this thesis proposes a twoway feature fusion method based on deep convolutional neural networks.This method enhances the feature maps for object detection by fully fusing the features of different layers,thus to improve the performance of object detection.After obtaining the feature maps of different scales,the method transfers the detailed features from the low-level layer to the high-level layer and the semantic features from the high-level layer to the lowlevel layer simultaneously and fuses these features obtained by transferring.Finally,the generated feature maps contain both the detailed features and the rich semantic features.Experimental results show that the proposed method improves the performance of object detection,especially for small objects,by fusing the features from different layers.2.Aiming at the problem that the visual features of small objects are not evident in complex scenes,this thesis proposes a novel object detection method based on context learning.This method improves small object detection by fully mining the context relationship between the objects and the background scenes in images.The proposed network consists of two parts,one is a three-layer perceptron and the other is a two-layer convolutional neural network.The three-layer perceptron is mainly used to learn the context relationship between object pairs in the image,and the two-layer convolutional neural network is used to learn the global context,that is,the relationship between the object and the entire scene.Experimental results show that the proposed method improves the performance of small object detection in complex scenes by making full use of context information.3.To reduce the missing rate of small objects caused by the anchor mechanism,this thesis proposes a novel small object detection method based on center and size prediction,which transforms the object detection task into an object center prediction problem and a size regression problem.Different from the existing methods with the anchor mechanism,the proposed method detects objects by center prediction and size regression.In our method,an object is modeled as a cross composed of the center and the size.Specifically,we propose a novel cascaded center prediction method that introduces a coarse-to-fine idea to improve the center prediction.Furthermore,since center prediction(considered as a classification task)is easier than size regression,we propose a center-attention size regression module that uses the output of the center detection to assist size prediction.Experimental results show that the proposed method can greatly reduce the missing rate of small object detection.