The System for Eliminating Blurring of Text Images Using Convolutional Neural Networks
Nataliya Shakhovska, Andrii Natiahlyi · 2021 IEEE 16th International Conference on Computer Sciences and Information Technologies (CSIT) · 2021
Image blur degrades the quality of human perception of images and significantly complicates their subsequent analysis using computer vision systems. Real blurs tend to have unknown, spatially complex, and unique blur kernels and usually are even more complicated by noise and other artifacts. The above factors determine the relevance of developing, analyzing, and improving deblurring methods based on convolutional neural networks. A thorough study of existing solutions in the field of machine learning and convolutional neural networks for restoring blurry images was conducted. This work provides a detailed description of the key elements of convolutional neural networks, describes the loss function used for training image recovery models, and provides an overview of perception loss. As a result of the thesis, a new architecture of deep convolutional neural network called CNN was proposed, which, while having 25 times fewer parameters, achieves similar results to the CNN-L15 network on PSNR, SSIM metrics, and during visual comparison of resulting images. Also, modifications to the CNN-L15 network architecture and usage of perceptual loss function are proposed, which help the network achieve significantly better results on metrics, and subjective visual clarity of images, while being trained the same number of epochs