Adaptive compression algorithm based on JPEG2000 with neural network corrector of decoded image definition

The article describes features of digital image processing in the process of adaptive compression based on the discrete wavelet transform (DWT) used in the JPEG2000 compression standard. Unlike the well-known compression algorithm, the new algorithm uses a reduction (scaling) of the sizes of the matrices of the DWT coefficients of the Y, U, and V signals at the first DWT iteration, after which a reduced copy of the low-frequency range image is used for further processing. Thus, at this stage the volume of video data is reduced by four times. To obtain the original resolution, in contrast to the well-known interpolation methods, the inverse DWT transform with added zero coefficients in the high-frequency subranges of the first iteration is used. The proposed algorithm allows compressing image files on average by 40...50 times with satisfactory quality. To restore high image quality, an original neural network sharpness corrector based on a convolutional model with training of sample image blocks by the brightness index in the Lab space is developed. Based on the ResNet architecture, a proprietary deep neural network model is developed, based on combining several techniques used to solve image reconstruction problems in other architectures. An optimal neural network training option is selected, allowing the trained model to be used to correct the clarity of decoded images to high objective quality indicators with subjective assessments of "good" and "excellent".