Research On Multi-Scale Object Detection Based On Foreground Features

In recent years,with the continuous enrichment of data sets in the field of object detection and the continuous optimization of network structure,the accuracy of object detection methods has been continuously improved.Compared with the original image,the image detected by the object detection method has a clear and significant candidate box,which makes the latter program easier to operate.How to mine the feature information in pictures to a greater extent and detect objects of different scales has gradually become the focus of widespread concern.At this stage,object detection usually carries on the feature extraction to all the positions in the picture,and then carries on the output prediction after the dimension splicing of the extracted features.This method ignores the interference of background information in the prediction process,and ignores the different contribution of different layers to the output results.For the phenomenon that there is a large background interference and a large change in the foreground scale in the output image,it is easy to have a high error detection.In view of this phenomenon,this paper studies the multi-scale object detection method based on foreground features.The main contents are as follows:(1)Research on the method of foreground feature extraction based on deformable convolution and attention mechanism.In view of the irregular shape of most foreground in the image,the irregular sampling point migration method based on deformable convolution is analyzed,and the network structure of rough feature extraction based on deformable convolution is studied.The foreground instead of the whole image is extracted to obtain more sufficient foreground feature.On this basis,considering the interference of background elements in the process of feature extraction,spatial and channel attention is adopted to study the network structure of fine feature extraction based on attention mechanism,so as to obtain more sufficient foreground features.(2)Research on multi-scale object detection method based on weighted fusion of foreground features.In view of the fact that the feature fusion network only carries on the dimensional splicing of the upper and lower layer information,ignoring that different layer features have different contribution in the fusion,the weighted fusion method of foreground features is analyzed,and the network structure of weighted fusion of foreground features is studied.At the same time,shortcut and multi-layer output are introduced into the network,the loss function of multi-scale object detection is analyzed,and the multi-scale object detection method based on feature weighted fusion is studied to improve the effect of multi-scale object detection.(3)Verification and implementation of object detection prototype system.Based on COCO and VOC object detection data set,combined with multi-scale object detection,a object detection system is designed and developed.On this basis,the error correction system is designed and optimized by model tuning based on genetic algorithm.The experimental results show that the proposed multi-scale object detection method can not only ensure the model size and computation,but also achieve a higher level of accuracy and speed,and better complete the object detection tasks in daily life.