Research On Deep Learning Techniques Based On Deep Residual Network And Its Applications In Vision

With the continuous development of machine learning technology,deep learning is widely used in various fields and can help make breakthroughs.The most widely used area of deep learning is computer vision,and the most commonly used deep discriminant learning model in computer vision is deep residual network.Deep residual network can alleviate the problems of gradient disappearance and network degradation which are difficult to overcome in the training of general deep network by adding parallel skip connection on the original weight layer.The introduction of this residual core makes the deep network deeper and deeper,the features that can be extracted more and more advanced,and the performance of the network better and better.On this basis,this paper conducts innovative deep learning technology research on deep residual network and applies it to two popular applications in computer vision research fields,namely pedestrian re-identification and medical image synthesis.The main research work of this paper is as follows:1.A full-scale deep metric learning method is proposed by combining a deep residual network considering depth,width,cardinality and multi-channel with a new metric learning method.It is applied to the problem of pedestrian re-identification to realize the matching task of pedestrian recognition in images taken by multi-camera systems with non-overlapping field of view.Comparisons were made with other 9 different metric learning methods on the CUHK01,VIPeR and QMUL-iLIDS pedestrian datasets.The experimental results show that the full-scale deep metric learning method is superior to other comparison methods in balanced situations and ranks third among all comparison methods in non-balanced situations.Overall,high performance can be achieved.2.An unbalanced multi-channel model is formed by combining residual networks of different depths,so that different hierarchical features learned from residual networks of different depths can be fused.At the same time,a new sub-structure based on residual network and three different sub-structure building blocks(input block,basic block and bottleneck block)are introduced to improve the generalization ability of the model as a whole.On this basis,an unbalanced deep discriminant learning model is constructed and applied to arterial spin labeling image synthesis in medical images.The 3 other deep learning models were compared on an Alzheimer's disease data set of 355 patients.The experimental results show that the arterial spin labeling images synthesized by the unbalanced deep discriminant learning model are closer to the gold standard than other models.At the same time,the average accuracy of Alzheimer's disease diagnosis can be significantly improved by combining the arterial spin labeling image synthesized by the unbalanced deep discriminant learning model and the real structural magnetic resonance image,reaching 62.48%(reference value 51.20%,gold standard 66.14%).3.A new UA-GAN ensemble based on U-net(including deep residual network)and residual attention component is proposd.Different networks in this ensemble will focus on different regions of medical images,and use flow-based Glow model to generate noise(features)based on the Gaussian mixed model to better cope with the common heterogeneity features in medical images,making the synthesized arterial spin labeling images more excellent and can use the same model to synthesize structural magnetic resonance images,that is,bi-directional synthesis.On the basis of two Alzheimer's disease data sets,extensive experiments were conducted on the proposed UA-GAN ensemble.The superiority of UA-GAN ensemble is confirmed compared with 7 different deep learning models.It also significantly improved the average accuracy of Alzheimer's disease diagnosis,reaching 73.71%(based on synthetic arterial spin labeling images,baseline 72.12%,gold standard 75.94%)and 72.73%(based on synthetic structural magnetic resonance images,baseline 71.80%,gold standard 75.94%).Meanwhile,for the first time,arterial spin labeling images were successfully synthesized from structural magnetic resonance images in the ADNI-1 data set.Through the research of three different innovative deep learning technologies based on deep residual network,it can be seen that the combination of deep residual network and different deep learning or metric learning technologies has great advantages.