Modeling statistical regularities is the problem of representing the pixel distributions in natural images, and usually applied to solve the ill-posed image processing problems. In this paper, we present an extremely efficient CNN architecture for modeling statistical regularities. Our method is based on the observation that, by random sampling the pixels in natural images, we can obtain a set of pixel ensembles in which the pixel value is independent identically distributed. This leads to the idea of using 1*1 (point-wise) convolution kernel instead of k*k convolution kernel to learn the feature representation efficiently. Accordingly, we design a novel architecture with fully point-wise convolutions to greatly reduce the model complexity while maintaining the representation ability. Experiments on three applications: color constancy, image dehazing and underwater image enhancement demonstrate the superior performance of our proposed network over the existing architectures, i.e., using 1/10-1/100 network parameters and computational cost over the state-of-the-art networks while achieving comparable accuracy. Codes and models will be made publicly available.

本文提出了一种基于深度学习的CNN网络结构来模拟自然图像的统计规律，并使用随机重排像素的方式提高CNN对于像素分布的关注度，采用点卷积核取代大卷积核，实现了更好的统计表示能力。实验结果表明，该网络在颜色校正和图像去雾两个应用中达到了可比较的性能，并且参数数量和计算成本比现有架构更小。

用全点卷积神经网络对自然图像的统计规律进行建模