| Human attributes contain a lot of important information,including appearance features such as people’s gender,age,clothing,etc.Human attribute detection is a challenging problem in computer vision research and is the basis of many computer vision research,such as pedestrian tracking,human weight recognition and Smart monitoring,etc.In the past few years,with the rapid development of deep learning,the application of deep learning algorithms to human attribute detection has become a new research hotspot,and good results have been achieved.In the task of human attribute recognition in surveillance scenes,the captured images are degraded due to camera pixels,movement of people,etc.The image degradation is often manifested as blurred images,resulting in the loss of main feature information,resulting in insufficient feature extraction in existing related algorithms.,The problem of unsatisfactory identification accuracy.In addition,there are few existing human attribute detection algorithms that focus on blurring image data.In view of the above problems,this thesis proposes a human attribute detection algorithm based on image restoration.The main research contents are as follows:(1)Image restoration-image Deblurring algorithm.Based on the existing DeblurGAN+network structure,the sparse gradient is prevented by replacing the activation function;the convolutional attention mechanism(CBAM)based on channel attention and spatial attention is added to the residual block of the original network structure to enhance The ability to extract specific features;improve the loss function,propose the use of richer content loss,and improve the model’s ability to deal with outliers.After the improvements in the above three aspects,the image deblurring algorithm proposed in this paper achieves 29.93 and 0.949 respectively in PSNR and SSIM in the GoPro data set,and the image restoration effect is superior to other comparison algorithms in both subjective and objective evaluation.(2)Human attribute detection algorithm based on improved ResNeSt network.Firstly,starting from the ability to extract features at multiple scales,pyramid convolution is integrated into the split attention network to capture different levels of details;for uneven sample distribution between attribute categories,uneven distribution of difficult samples and uneven distribution of attribute categories The problem is that dynamic weights are introduced into the loss function to dynamically adjust the weights of positive and negative samples to improve the detection effect of the model.Experiments show that the human attribute detection algorithm of the improved ResNeSt network achieves 74.57%and 78.34%accuracy in the two evaluation standards of Accuracy and mAP in the original PA-100K data set.Accuracy and mAP on the two evaluation criteria have reached 77.62%and 81.84%accuracy,which are better than the baseline model and other existing comparison models. |
