Research On Small Object Detection Based On FSSD

The task of object detection is to identify the category of objects in images or videos and locate their positions.It is a basic research in the field of computer vision,and it is widely used in some industrial and life fields.Compared with traditional methods,methods based on deep learning can extract more complex features using convolutional neural networks in object detection,and greatly improve the speed and accuracy.However,there are still problems such as error detection and missed detection,especially It is still a difficult problem to detect small targets(the target size is 10% of the original image area,or the target size is less than 32×32 pixels in the MS COCO data set).This paper proposes two improved methods based on the deep learning algorithm FSSD(Feature Fusion Single Shot Multibox Detector).They are adding an improved feature fusion network and an improved method of extracting features using a de novo training network to assist the pre-training model.We are in the object detection general data Integrate MS COCO and Pascal VOC for verification,and at the same time do further testing on the strip steel surface defect detection data where small targets occupy a large proportion.The main research work is divided into the following two parts:(1)This article analyzes the advantages and disadvantages of FSSD.Aiming at the problem of low detection accuracy of small targets,a shallow feature map extracted from the backbone network is added to the Receptive Field Block(RFB),and two feature pyramids are added at the same time.Feature Pyramid Networks(FPN)is a method of fusing multi-scale feature maps.In short,this method is based on the idea of FSSD and integrated into the idea of FPN to predict the target.Due to the use of the same backbone network,the increased amount of calculation has not been significantly improved,so there is no significant loss in detection speed.In the Pascal VOC data set,the detection accuracy of small objects in bottles and pots has been improved;for the MS COCO data set,the detection accuracy of small objects has increased by 2.6%;in the hot-rolled steel strip surface defect data set(NEU-NET),the recognition accuracy of small and elongated defects such as pits and scratches are improved by 2.2% and 4.6%,respectively;Mean Average Precision(m AP)on the self-made surface defect data set reaches 58.4%,The detection speed reaches 51 frames.(2)In the object detection algorithm,the mainstream method is to use the pre-trained classification model as the backbone network to extract features,but it focuses more on the classification task in the detection process,which leads to deviations during algorithm training,which is not conducive to target positioning;in addition,pre-training The structure of the model is fixed,and the frame modification is not convenient.Based on these problems,this paper proposes another improved method for the FSSD algorithm.It uses the de novo training network to assist the pre-training classification model to extract features,complements the features extracted by the two methods,and then adds the feature pyramid using the feature map channel fusion method to combine the deep features The rich semantic information is propagated to the shallow features,and the shallow feature maps used for prediction have rich texture and semantic information at the same time,which is conducive to the detection of small targets.In addition,by reducing the number of channels of the original feature fusion module of FSSD,unnecessary calculation costs are reduced.In the Pascal VOC data set,the detection accuracy of small objects in bottles and pots has been improved;in the MS COCO data set,the detection accuracy of small objects has increased by 2.6%;in the NEU-NET,pitting and scratches,etc.The recognition accuracy of micro and slender defects was increased by 0.3% and 4.1% respectively;on the self-made surface defect data set,the m AP value reached 55.5%,and the detection speed reached 52.5 frames.