Research On Small Object Detection Based On ResNeSt Framework

Computer vision aims to recognize and understand content in images or videos.It has four basic tasks of localization,classification,detection and segmentation.The purpose of object detection consists of two subtasks: classification and bounding box regression.However,a large number of high-quality deep learning frameworks have been proposed and open sourced to improve the performance of general object detection as many researchers poured into the field of object detection.This provides a new solution for object detection.When the deep learning method is applied to object detection,the feature information with stronger feature expression ability can be learned through the neural network,which greatly improves the detection performance of the object detection algorithm.When the object detection can be expanded to accurately detect the very small object in the image,it is the small object detection.As a derivative neighborhood of object detection,it has been a difficult point that has not been effectively overcome for a long time.Because it can play an important role in many fields such as military field,autonomous driving,face payment,medical field,Skynet system,human-computer interaction,urban monitoring and intelligent transportation,it has a wide range of application prospects.Small objects usually represent a very small proportion of the entire image or video,and the appearance information is relatively scarce,so it is difficult to detect them from similar objects or cluttered backgrounds.In addition,the application scenarios of small target detection in real-life production are intricate and often accompanied by problems such as target occlusion,background clutter,target scale changes,background blurring,and drastic changes in illumination.It seriously affects the feature extraction of small targets,and further increases the difficulty of detecting small targets.Although the current research on object detection can be put into practical application,due to the lack of pixel information of small objects in images and the lack of dedicated datasets in this field,the current research on small object detection still needs to be done.Among the currently proposed methods to improve the accuracy of small target detection,when the resolution of the input image is increased,the pixel information of the small target is increased,but at the same time,the consumption of computing resources is increased,resulting in a decrease in detection efficiency;in addition,the use of multi-scale features Fusion methods to improve the detection accuracy of small objects cannot guarantee that the constructed features are distinguishable and interactive.Similarly,various techniques of deep learning are used to integrate into small target detection.Although the network can be deepened to extract more representative and abstract features,the location information of small targets will also be lost.At the same time,we found that there is a big gap in the significant detection performance of many excellent research results.It can be seen that the research of small target detection is still a research field that attracts researchers and is full of challenges.Because the research on small target detection has its important research significance and practical application value,this paper is devoted to optimizing the feature extraction module of small target and improving the algorithm of comprehensive performance of small target detection.The following solutions are proposed:1)This paper proposes a Faster Rcnn small target detection FRRSnet network model based on feature fusion,which fuses the deep features and shallow features extracted by the Res Ne St feature extraction module in the Faster Rcnn network to obtain information with internal data structure and spatial hierarchy.strong features.In addition,the method of hole convolution is introduced in the convolution layer,so that the receptive field can be improved under limited computing resources without reducing the resolution of small targets,thereby enhancing the feature expression ability of the Faster Rcnn network model,thereby improving the algorithm’s ability to respond to small objects.Object detection accuracy.2)This paper also proposes the Faster Rcnn network based on the convolutional attention mechanism,referred to as the FRcbam network.The convolutional attention mechanism includes the spatial attention mechanism and the channel attention mechanism.The spatial convolutional attention mechanism can retain more information related to target positioning and improve the positioning accuracy of objects.The channel convolutional attention mechanism can According to the relevant feature information of the object,the classification accuracy of the target is improved.3)In the experimental part,this paper uses the general data set PASCAL VOC and the traffic sign detection data set Tsinghua-Tencent100 K for training to evaluate the generalization ability of the model.Verification,and use the popular small target detection algorithm for algorithm comparison,through a large number of real-time results analysis,it shows that the algorithm proposed in this paper has achieved excellent detection performance on the basis of the benchmark Faster RCNN.