ENCODING RASTER IMAGES BASED ON FRAGMENT SIMILARITY

Abstract

This paper is devoted to image encoding based on determining the similarity of fragments by using neural networks to extract the features of fragments and machine learning algorithms to find similar fragments. In the modern world, the problem of image storage is quite relevant. Graphic data takes up quite a lot of disk space, while Internet users upload more and more pictures. Also, every year there is a development of photography and image quality is improving, respectively, and the size of graphic data is growing. Data warehouses of social networks, messengers, file sharers and other Internet resources are filled with tens of thousands of new pictures every day. Therefore, the question arises about reducing the size of graphic data. In general, it should be noted that one of the most important and defining aspects of both storage and transmission of information is its compression. The problem described above is solved by encoding and compressing images. With the help of coding, the size of graphic information is reduced, which saves storage space and, accordingly, the money spent. In view of this, it is important to develop a method and means of image coding. Many methods exist for compressing graphic information. For example, jpeg, webp, png and others. These methods usually use the removal of redundant information in the photo and work purely with the image itself, but none of the methods uses fragments of similar images. The article uses convolutional neural networks and KNN (k-Nearest Neighbors) classifier for image encoding. and compares the size of the encoded image with the input. In order to encode the image, you first need to fill the data warehouse with features of fragments of similar images, then for each fragment of the obtained images you need to select the features and write to the data warehouse. Once the snippet feature database is formed, you can encode new images using saved snippets.