【图像质量评价】Convolutional Neural Networks for No-Reference Image Quality Assessment ---2014
本文主要探讨了图像去模糊技术的相关内容。首先介绍了 NSS 基于的特征方法及其局限性,并提到 Cornia 方法可以直接从原始像素中学习判别性图像特征。其次提出了 ResNet-50 网络结构,并通过 Local contrast normalization 对图像进行标准化处理。此外还讨论了网络中的池化操作(3x3 最大池化)和激活函数(Relu),并采用变动量 SGD 优化算法以提高训练效果。最后对 LIVE 和 TID2008 数据集进行了实验验证,并提出了基于重叠采样的改进方法以平衡样本数量与采样效率的关系。
一、直接引用
Significant achievements in computer vision rely on the use of NSS-based feature descriptors.
Usually, NSS-based features are utilized to describe the distributions of particular filter responses. Traditional NSS-based features are typically extracted within image transformation domains, such as those employing the wavelet transform [10] or discrete cosine transform (DCT) [13]. These approaches generally exhibit low efficiency due to their reliance on computationally intensive image transformations.
Experimental results demonstrate that CORNIA enables the learning of discriminant image features directly from raw pixel data, rather than relying on handcrafted features.
With equipping a deep structure,the convolutional neural network (CNN) has the ability to effectively learn complex mappings without requiring extensive domain knowledge.
Increasing the depth of the network will substantially enhance its learning capacity.
P4**:** As a result of applying dropout to all layers, the time required for model convergence is significantly prolonged. Therefore, in our implementation, we limit applications of dropout solely to the second fully connected layer.
二、作者的观点
通过应用局部对比度归一化技术后对图像进行标准化处理,并随后执行裁剪操作
(2)此文献裁剪的patch为32*32 , 5层网络;
采用了SGD优化算法作为训练方法,并引入了一个可变的动量项moment。该动量项在其值在前期epoch期间保持在较高的水平(如保持不变于值为 ),随后降至较低水平(如降到 )。
(4)LIVE 数据集: Higher DMOS indicates lower quality,
TID2008数据集: Contrary to DMOS, higher MOS indicates higher quality。
基于非重叠取样的方法,在获得的样本数量有限的情况下(即当样本数据较为稀少时),需要采取重叠采样的方式以确保数据完整性。(当样本数据较为丰富时(即在存在充足的数据资源情况下),推荐优先选择非重叠取样的策略)
网络结构 为:

采用了灰度图像作为输入,并选用Relu作为激活函数。值得注意的是,在网络结构中仅在最后两个全连接层应用了min pooling方法。
三、我的观点
(可以借鉴局部归一化 )
(使用最大最小池化,则最小池化前不能加Relu)
(SGD 使用变momentum)
