| The construction of public security video surveillance network is an important means to safeguard national security and social stability,prevent and combat violent terrorist crimes in new situation.Due to the influence of various internal and external environment,different types of distortion will be introduced to surveillance videos.Traditional blind assessment algorithm is mainly for a specific distortion of static image,while we need to consider more complicated scenario content,and realize blind video quality assessment under the condition of the unknown distortion type.Therefore,to solve the blind quality assessment problems of blind distortion video,this paper has done some related research work combined with pattern recognition and artificial intelligence technology.To slove the blind discriminant problem of distortion surveillance video,this paper proposed a recognition method for blind distortion type videos based on CNN.We segment the static images which derived from sample videos into small patches as CNN's input.The distortion type of image blocks are predicted by the trained CNN and the majority voting rule are adopted to decide the class of videos.To reduce overfitting and local minimum,the positive and negative sample equalization and an adaptive learning rate were introduced.For the problem of blind video quality assessment,this paper proposed a hybrid learning model based on Convolution Neural Network.The problem of predicting video image quality is formulated as a classification problem by converting the benchmark scores into integers as class labels to train a CNN.The output of the last full connection layers are extracted as deep features,along with the corresponding benchmark scores are used to train a SVR,and the quality score can be predicted by the trained SVR.We introduced the global pooling instead of traditional local pooling method,which greatly reduce the feature dimension and better to extract the statistical features of image.In order to explore the detection problem of leaf occlusion video with unlabeled data,a leaf occlusion detection method based on deep neural network is proposed in this paper,the result is comparatively satisfied. |
