Single-stage Target Detection Method And Application Based On Deep Learning

In computer vision technology,the one of the important technologies is target detection,which can obtain the types and positions of objects.Among them,the single-stage target detection methods have simple structure and high computational efficiency.They are widely used detection methods.Because the existing single-stage target detection algorithms have the problems of weak feature extraction ability and low efficiency of feature fusion in the face of different scale images,the detection accuracy results of the algorithms are low.This paper deeply studies the relevant ideas such as weighted Chebyshev distance,feature pyramid and class balance.At the same time,it also deeply studies the existing methods of feature extraction and feature fusion in target detection.The main research results are as follows:(1)An image segmentation method based on weighted Chebyshev distance is proposed.Firstly,the image gradient information and neighborhood information are combined,the noises are removed by threshold method,and the feature vectors are extracted from the RGB spatial pixel matrix of the corresponding images;Secondly,the weighted Chebyshev distance is calculated to obtain the similarity matrix,and the image segmentation is realized by the spectral clustering;Experiments show that this method can describe image features more truly and efficiently.Compared with Otsu algorithm,this method’s segmentation speed is about 9 times faster,the running time is reduced by 523.2 seconds,and its accuracy is about half higher than MDA algorithm.(2)A single-stage target detection method based on loss function and feature fusion is proposed.Firstly,the inherent length and width constraints of the existing bounding box loss functions are transformed into diagonal constraints,and this chapter proposes the SIOU loss function;Then,the mixed hole convolution module and deconvolution pool module are integrated into the existing feature pyramid structures,and the feature fusion network De PANet is designed;Combined with the above contents,a single-stage target detection algorithm based on SIOU loss function and De PANet feature fusion network is proposed.Finally,experiments show that its detection AP value is increased by about 9% and its detection time is shortened by about 10 seconds compared with the original YOLOV4 algorithm.(3)A target detection prototype system for galaxy pair detection is designed and implemented.The system mainly includes: data preprocessing,sample labeling,generation of detection training model,result statistics and other modules,which are realized through several steps: data preprocessing,label generation,model training and so on.The results show that the system provides an effective way to further obtain galaxy pair candidates.