Research On Feature Extraction Methods For Small Object Detection Based On Deep Learning

The convolutional neural networks(CNNs)make object detection technology develop rapidly.Small object detection is an important part of the object detection tasks,and it is also an urgent problem to be solved at present.Because the area occupied by the small object in the whole image is relatively small,and with the deepening of the number of network layers,the resolution and semantic information of small object features will become less and less obvious.In the later stage of feature extraction,the information of the small object is almost difficult to be extracted by the network.For such reasons,the performance of small object detection is far less than that of generic object detection,so it has also attracted the attention of an increasing number of experts and scholars.In the structure of object detection networks,there is no doubt that the backbone network for feature extraction is an indispensable part of the whole detection network.Therefore,to optimize the performance of small object detection tasks,this paper mainly makes some improvements.(1)This paper presents a new feature extraction strategy SAFF-Net based on the attention mechanism.According to the characteristics of small object resolution and semantic information in different feature layers,a sandwich attention feature fusion module(SAFF)module is proposed.The module consists of two channels of attention and one spatial attention alternately.It enables the network to focus more on the characteristics of small objects,facilitating the preservation of the shallow details of small objects and the enhancement of highlevel semantic information.The SAFF module can combine any feature extraction neural network to obtain the weight data of attention.To verify the performance of network feature extraction,image classification experiments and object detection experiments are carried out respectively.The classification experiments added the SAFF module to the 50-layer residual network(ResNet-50)and 101-layer residual network(ResNet-101)respectively,which formed a new classification network SAFF-Net-58 and SAFF-Net-109.The Corona Virus Disease 2019(COVID-19)dataset was used in this study.The results show that the method improves the average accuracy of COVID-19 to 98.163%.In the object detection experiments,the SAFF module is added after each stage in the feature extraction process of Mask R-CNN.Taking Microsoft Common Objects in Context(MS COCO)as the experimental object,the results show that this method improves the average precious(AP)of small object detection by 4.0%based on Mask R-CNN.(2)This paper presents a small object detection feature extraction network DSAFF-Net based on dilated convolution.This method has two changes.One is to replace the pooling operation of C1 obtained in the original feature extraction network by using dilated convolution and point convolution operations,the other is to use expansion convolution technology to construct a new feature extraction module,the D-block module,to replace the method of obtaining P6 feature layer in the original network structure.These changes together form the new detection network DSAFF-Net.The net is used to alleviate the drawback of partial feature loss and reduce the spatial resolution due to down-sampling to reduce the feature map to expand the perceptual field and to fully improve the feature extraction capability in small object detection.In experiments using the MS COCO,the results show that the detection precious of small objects is improved by 4.4%compared with Mask R-CNN.Compared with SAFF-Net,the detection precious of small targets is improved by 0.4%.