The Application And Research Of Data Augmentation And Model Compression On Person Re-Identification

Person re-identification(re-id)refers to the technique of using computer vision techniques to determine if a matching person is appearing in disjoint camera areas.Person re-id has attracted the attention of industry and academia because of its potential application value in the video surveillance and criminal investigation.The main challenge of person re-id lies in three points:(1)there are strikingly discrepancy between different camera views,including illumination,background and human pose(2)the resolution of the data set is low and only pedestrians can be used Coarse-grained features such as contours and colors,and cannot use the fine-grained features commonly used in image recognition tasks,such as facial features(3)Current person re-id methods is highly dependent on deep neural networks therefore highperformance computer hardware is required.In order to alleviate the problem brought by the shift of camera view and the dependence of current pedestrian re-recognition methods on complex models,this thesis consists of two parts: The first is to use the advantages of the generative adversarial networks in the field of image generation to mitigate the impact brought by different camera views,thereby improving the accuracy of person re-id.The second is to embed the model compression technology into the current person re-id model,which greatly reduces the model complexity without sacrificing too much recognition accuracy.Existing person re-id methods rely mostly on implicit solutions,such as seeking robust features or designing discriminative distance metrics.Compared to these methods,human solutions are more straightforward.That is,imagine the appearance of the target person under different camera views before matching target person.The key idea is that human can intuitively implement viewpoint transfer,noting the association of the target person under different camera views but the machine failed.In this paper,we attempt to imitate such human behavior that transfer person image to certain camera views before matching.In practice,we propose a conditional transfer network(c Trans Net)that conditionally implement viewpoint transfer,which transfers image to the viewpoint with the biggest domain gap through a variant of Generative Adversarial Networks(GANs).After that,we obtain hybrid person representation by fusing the feature of original image with the transferred image then perform similarity ranking according to cosine distance.Compared with former methods,we propose a humanlike approach and obtains consistent improvement of the rank-1 precision over the baseline in Market-1501,Duke MTMC-Re ID and MSMT17 dataset by 3%,4%,4%,respectively.(2)Although the existing deep learning network can achieve a high recognition accuracy for most person re-id datasets,the structure of most network is complicated and the calculation amount for network is tremendous.This also directly leads to the dependence on hardware devices such as memory and powerful GPUs.Based on the above,we embed the pruning technology into the existing person re-id model.In experiments,we used Des Net pre-trained on Image Net as the baseline model,and then fine-tuned on the Market-1501 dataset.According to the experimental results,we proved that on Market-1501,our method can maintain a relative high accuracy,and at the same time can reduce the model parameters by 58% and reduce the amount of calculation by 57.12%.