Lі F. Design and analysis of efficient methods for providing a desired quality in image lossy compression

Lossy image compression has been developed to be an essential tool in the past several decades. This is due to the rapid development and broad application of imaging technology that has resulted in a sharp increase in the number of images and an increase in images size. Compared to lossless compression, lossy compression can achieve a higher compression ratio. However, inevitable distortions are introduced, which determine the visual quality of decompressed images. Consequently, the visual quality needs to be evaluated and distortions need to be controlled in practical applications. The thesis is devoted to solving the scientific and applied problem of increasing the distortion controlling efficiency in lossy compression with controlling parameter determination. The object of the study is the controlling of image distortions introduced by lossy compression. The subject of the study is the method of providing a desired visual quality in lossy compression. An analytical review of existing popular lossy compression coder’s performance in terms of compression ratio and several visual quality metrics, purpose and requirements of visual quality controlling in typical applications, as well as types and characteristics of different methods of distortion control are performed in the thesis. It is clear that the visual quality of decompressed images impacts further image processing and should be controlled carefully. The trade-off between distortion and compression ratio is often looked for in lossy compression, and compression control parameters are calculated according to this purpose. Based on the various requirements of users, the perceptual lossless effect can be demanded, and more flexible desired visual qualities could be set by users. It has been shown that the visual quality of decompressed images depends not only on the compression control parameter but also on the complexity of an original image, and the encoder adopted. The main problem of existing image lossy compression methods, which take account into visual quality, is either unsatisfactory accuracy with an acceptable time computation or appropriate accuracy but inappropriate or uncertain time efficiency. In the thesis, the actual scientific and applied task of designing lossy compression method for providing a desired visual quality, which considers control accuracy of different coders combined with analysis of the method efficiency is set and solved. The method of predicting visual quality of decompressed image, the two-step lossy compression method and the improved adaptive method which groups images based on their complexity, the method of multi-channel image lossy compression have been designed and analyzed with several visual quality metrics. The thesis aims to design efficient method of lossy compression to provide a desired visual quality, and analyze the accuracy of distortion control, in particular the lossy compression method for the multi-channel image in remote sensing, and to analyze the effect of visual quality on the accuracy of classification. The scientific results are: 1) For the first time, a two-step method is proposed to provide the required visual quality in gray-scale lossy compression. The average rate-distortion curve obtained offline is used to calculate the initial parameters, and the visual quality in the first-step compression is corrected by feedback. The results show that, in terms of human visual system (HVS)-based metrics, the accuracy of the second-step compression of general images is good enough, and the residual error is acceptable. 2) The method of predicting the visual quality of decompressed images for given parameters has been improved and extended to lossy compression encoders based on DWT. 3) The primary two-step method has been modified to enhance its robustness, and two methods are used. The first is to correct the constraints of the scheme to reduce the errors for high-texture images; the second is to pre-classify images according to their complexity and then use adaptively the proper average rate-distortion curve. 4) The two-step method has been extended to color and three-channel images, and its application in remote sensing has been discussed. The results show that the classification accuracy of compressed images is roughly the same as for the original (uncompressed) data if there are no visually noticeable distortions. In addition, it is possible to control the quality of 3D compression. The compression ratio it produces is twice that of wise-component compression, and the probability of correct classification is slightly higher. The study has been carried out for DCT-based coders (including AGU and ADCTC), DWT-based SPIHT coder and novel BPG coder, and the universality of the method for different images has been discussed.