Abstract—Image capturing is often performed with distortions caused by inaccurate focusing, displacement of the scene or recorder, radiation dispersion in the signal transmission medium, and similar aberrations. The distortion operator, which is commonly unknown, is needed for image restoration. Therefore, the determination of type and parameters of distortions using the observed signal is actual. In the paper a new approach to distortions diagnostics of video information by means of deep neural networks is proposed. The task of determination of the type and parameters of the main linear homogeneous distortion operators (circular with rectangular profile, circular with Gaussian profile, and linear with rectangular profile) is considered. An application of neural networks with the ResNet50, ResNet29, and ResNet18 architectures to identify the type and to determine the distortions parameters is studied. The research shows that the efficiency of the neural network is no less than that of methods based on direct spectral analysis.

摘要-图像捕获通常会因聚焦不准确，场景或记录器的位移，信号传输介质中的辐射色散以及类似像差而导致的变形而执行。图像恢复通常需要未知的失真算子。因此，使用观察到的信号来确定失真的类型和参数是实际的。本文提出了一种通过深度神经网络对视频信息进行失真诊断的新方法。考虑确定主要线性齐次畸变算子（具有矩形轮廓的圆形，具有高斯轮廓的圆形和具有矩形轮廓的线性）的类型和参数的任务。神经网络在ResNet50，ResNet29，研究了ResNet18体系结构以识别类型并确定失真参数。研究表明，神经网络的效率不低于基于直接频谱分析的方法。