激光与光电子学进展, 2019, 56 (11): 111003, 网络出版: 2019-06-13
卷积神经网络结合深度森林的无参考图像质量评价 下载: 974次
Quality Assessment Without Reference Images Based on Convolution Neural Network and Deep Forest
图像处理 无参考图像质量评价 卷积神经网络 回归森林 局部对比度归一化 image processing quality assessment without reference images convolutional neural network regression forest local contrast normalization
摘要
提出了一种新的卷积神经网络与深度回归森林结合的无参考图像质量评价方法。该方法对原始图像进行局部对比度归一化处理,采用卷积神经网络提取图像质量的判别特征,最后利用深度回归森林预测图像质量。该方法无须手工设计图像特征,简化了图像的预处理过程。较少的卷积层数有利于减少网络的训练时间。使用深度策略对回归森林进行集成,提高了单一森林的预测精度。基于LIVE数据库与TID2008数据库的实验结果表明,该方法能很好地预测图像质量,并具有良好的泛化性能与较高的准确率,尤其在JPEG2000压缩、高斯模糊和白噪声等3种失真类型上均表现出了良好的性能。
Abstract
This paper proposes a new quality assessment method without reference images based on the convolutional neural network (CNN) and deep regression forest. First, this method performs a local contrast normalization on the original images. Second, it subsequently uses CNN to extract the discriminant features of the image quality. Finally, it utilizes the deep regression forest to predict the image quality. The method does not require any manual features, which simplifies the process of image preprocessing. In addition, fewer convolution layers are beneficial to reduce the training time of the network. The application of deep strategy to integrate the regression forests improves the prediction accuracy of a single forest. On the LIVE and TID2008 databases, the experimental results show that the proposed method can predict the image quality well and has a good generalization performance with high accuracy. The proposed method achieves a state-of-the-art performance, especially in JPEG2000, Gaussian blur and white noise distortions.
陈寅栋, 李朝锋, 桑庆兵. 卷积神经网络结合深度森林的无参考图像质量评价[J]. 激光与光电子学进展, 2019, 56(11): 111003. Yindong Chen, Chaofeng Li, Qingbing Sang. Quality Assessment Without Reference Images Based on Convolution Neural Network and Deep Forest[J]. Laser & Optoelectronics Progress, 2019, 56(11): 111003.