Research On Blurred Image Recognition Based On Transfer Learning

In recent years, the applications of the face recognition and video surveillance have received great attention. Due to the influence of the object-camera relative motion, atmosphere turbulence and out of focus, a lot of blurred images are gotten. Most common artificial descriptors such as LBP and HOG are blur-sensitive. Although many blur-insensitive descriptors such as LPQ and moments have been proposed, these descriptors are based on centrally symmetric assumption. However, most blur in reality can't meet central symmetric, so the recognition ability of these descriptors drops rapidly.Firstly, this paper proposes to solve the blurred image recognition task from the perspective of transfer learning. This mechanism is that, the training set composed of labeled clear images is seen as source domain, and the testing set composed of unlabeled blurred images is seen as target domain. As common descriptors are not blur-invariant, the training samples and the testing samples have a huge difference in the feature space distribution. In this paper, transfer learning based on subspace alignment is applied to make two subspaces closer to make a robust classifier.Secondly, this paper proposes a transfer learning method based on metric learning. In this method, the source subspace are built using the LMDR metric learning method, which makes full use of the labels of clear images and enhances the expression ability of source subspace. This paper also proposes a new subspace alignment method.Besides, this paper proposes another transfer learning method based on low-rank decomposition, which use the average sparse ingredients after the multilevel low-rank decomposition as the new characteristics of expression of two domains. The subspace establishment and alignment are executed with the new feature space and to solve the multi-blurring image recognition task. Lastly, this paper study the blur-insensitive ability of Convolution Neural Network(CNN) features, and we extract the features using VGG-16 and use the SVM classifier.Extensive experiments on three types of datasets including the face dataset, texture dataset and scene dataset demonstrate the effectiveness of three methods. This paper discusses two circumstances that the blurred images in the target domain are blurred by single blurring type and multi-blurring types. The experimental results demonstrate that three kinds of transfer learning methods can achieve good results in all the blurred image recognition tasks. At the same time, we can get that the features extracted from CNN has more blur-insensitive ability compared with common descriptors from the experimental results.