一种改进的卷积神经网络的无参考JPEG2000图像质量评价方法

摘要

The existing image quality evaluation model for JPEG2000 compression image distortion upon evaluation is not very ideal.In view of this, a JPEG2000 compressed image quality evaluation method based on improved CNN framework is put forward.The model is consisted of one convolutional layer with 20 convolution kernels, one pooling layer with the max, medium and min pooling, one fully connected layer with 1200 ReLU units and one output node.Using the max, medium and min pool structure to extract the typical features in the image effectively.Experimental results show that the proposed method is consistent with the subjective score better in the LIVE library.%现有的图像质量评价模型对JPEG2000压缩图像的失真情况评价都不是很理想.针对这一问题,提出一种基于卷积神经网络的JPEG2000压缩图像质量评价方法.该模型由一层包含20个卷积核的卷积层,一层包含最大池、中值池和最小池的次采样层、一层采用1200个ReLU激活单元的全链接层和一个输出节点构成.采用最大、中值、最小三池联合的方法,可以有效提取图像的质量感知特征.在LIVE图像质量评价库JPEG2000压缩图像上的实验结果表明,该方法得到了比相关文献方法更好的主观感知一致性.