基于卷积神经网络的视频图像失真检测及分类*

为了检测不同失真类型的视频图像，实现对失真视频图像的分类处理，本文提出一种基于卷积神经网络的视频图像失真检测及分类方法。首先，将视频图像分割成较小的图像块作为输入，然后利用卷积神经网络主动学习特征，引入正负例均衡化和自适应学习速率减缓过拟合和局部最小值问题，由softmax分类器预测图像块的失真类型，最后采用多数表决规则，得到视频图像的预测类别。采用仿真标准图像库(LIVE)和实际监控视频库对本文方法进行性能测试，前者的总体分类准确率达到92.22%，后者的总体分类准确率达到92.86%。整体的分类准确率均高于已有的其他三种算法。引入正负例均衡化和自适应学习速率后，CNN的分类准确率得到明显提升。实验结果表明本文方法能主动学习图像质量特征，提高失真视频图像分类检测的准确率，通用于任意失真类型的视频图像分类检测，具有较强的鲁棒性和实用性。

Video Image Distortion Detection and Classification Based on Convolutional Neural Network

To detect different kind of distortion in video images, and process the distortion video images according to the category. A detection and classification method of distortion video images based on Convolutional Neural Network (CNN) is proposed. Firstly, we take the small patches which are segmented from the video images as input. Then use CNN to learn image features , and predict the class of the patches with a softmax classifier. To reduce overfitting and local minimum, we introduce positive and negative sample equalization and a adaptive learning rate. Finally, the majority voting rule is adopted to decide the class of video images. We conducted the performance evaluation for the proposed algorithm on the standard image database (LIVE) and the real-world surveillance video database. The total classification accuracy was up to 92.22% on the former, and 92.86 % on the latter. The overall classification accuracy are higher than the other three algorithms. CNN has better performance with positive and negative sample equalization and a adaptive learning rate. Experimental results show that the proposed method can learn image quality features by itself, and improve the classification accuracy. It has good robustness and practicability, and suitable for any kind of distortion video images.