Analysis of the Repair Effect of Deep Learning Algorithms on Abnormal Images

Given the common reasons for abnormal phenomena in images and the harm caused by abnormal images, this paper proposes an unsupervised feature learning algorithm for pixel prediction based on image content structure to address the issue of image missing. The designed image content structure encoder based on convolutional neural networks not only understands the content of the entire image, but also generates a reasonable assumption for the missing parts. During feature learning, it not only captures the appearance, but also captures the semantics of the visual structure. Therefore, the content of any image region can be generated reasonably. When training the image content structure encoder, $\mathbf{L} 2$ construction loss and adversarial loss are used. And it was validated on the experimental dataset in the deep learning Caffe framework. Experimental results have shown that the network model designed in this article can improve the quality of image restoration and achieve high efficiency.