| As applications in intelligent surveillance,telemedicine,and vlog become more common and necessary,there has been an increasing demand for high bit and spatial resolution images that can provide clear and accurate texture details in image processing.While upgrading the hardware of imaging systems enables the acquisition of images with higher bit and spatial resolution,this approach poses significant cost and technical challenges.Furthermore,it cannot reconstruct the low-resolution images into highresolution ones.Therefore,it is essential to conduct research on algorithms that can enhance both the bit and spatial resolution of images.The existing algorithms for enhancing the bit and spatial resolution of images face the following problems:(1)In terms of bit resolution,effectively eliminating false contours caused by the degradation of bit is the most prominent and challenging issue.However,the characteristics of these false contours have not been sufficiently studied,making it difficult for existing traditional algorithms to effectively eliminate these artifacts.(2)Deep learning-based algorithms for bit resolution exhibit better image reconstruction potential.However,the existing algorithms have not explored effective network structures and data processing methods to fully leverage this potential.(3)In terms of spatial resolution,the image blurring problem caused by resolution degradation has been effectively addressed with deep learning methods,but many existing network structures are too complex to be practically applied.(4)Simultaneous degradation of bit and spatial resolution can result in false contours and edge blurring in the image,and the blending degrees of false contours and edge blurring vary significantly in different image regions.In this dissertation,the aforementioned problems are addressed,and the following results are achieved:(1)A bit resolution enhancement algorithm based on false contour elimination filtering is proposed to address the problem of poor false contour elimination in traditional bit resolution enhancement algorithms.The false contour is an edge that does not exist originally,so the essential function of this algorithm is to eliminate the false edges.The algorithm includes the following core sub-techniques: First of all,false contour elimination is modeled as edge elimination filtering,and a false contour elimination filter is designed;Secondly,based on the mechanism of false contour generation,an appropriate threshold for filter weights is designed,so that the filter can selectively eliminate false contours and preserve real edges;And then,based on the numerical characteristics of the pixels around the false contour,a false contour region detection method is designed,which can improve the filtering efficiency and alleviate the problem of over-smoothing in the detail regions;Finally,two adaptive mechanisms,namely direct size matching and iterative size matching,are introduced in the filtering process to handle different sizes of false contours simultaneously.Experimental results demonstrate that the proposed algorithm,based on false contour elimination filtering,effectively suppresses false contours in low-bit images while enhancing bit resolution.(2)Regarding the lack of research on network structure and data processing in deep learning-based bit resolution enhancement algorithms,this dissertation proposes two algorithms: Bit Enhancement via CNN with Auto-encoder Structure(BE-AUTO)and Planned Sensor Distortion and CNN-based Bit Enhancement(PSD-CNN).BE-AUTO specifically focuses on improving the reconstruction quality of high-bit images from existing low-bit images by optimizing the network structure.Unlike related algorithms that solely rely on convolution for network construction,BE-AUTO incorporates a variational auto-encoder structure comprising convolution and deconvolution layers.By performing layer-by-layer feature extraction through convolution,better image features are obtained.Furthermore,deconvolution facilitates feature reconstruction,resulting in improved image quality.Additionally,BE-AUTO introduces global residual learning to effectively reduce estimation errors and color distortion caused by full dynamic range mapping in existing methods.PSD-CNN,on the other hand,is designed to enhance the quality of the reconstructed images by establishing a correlation between low-bit image generation and high-bit image reconstruction.PSD-CNN employs a pre-distortion function to spread false contours to neighboring pixels,making them less perceptible.By leveraging multiple descriptions constructed through pre-distortion,and employing a suitable quantification method to avoid blurring,PSD-CNN achieves the generation of high-quality high-bit images.(3)To achieve better results,more and more complex networks are used in most of the existing spatial resolution enhancement algorithms,but the difficulty in training and limited applicable scenarios caused by the bloated structures and sizes of the networks should not be ignored.In light of this,this dissertation proposes an Image Superresolution Method Based on Lightweight Concatenated Residual Network with Channel Attention(LCRCA).The key objective is to maintain processing quality while keeping the network lightweight through more efficient feature extraction and utilization.LCRCA incorporates several techniques to achieve this goal.Firstly,it introduces cross-layer information concatenation to enable the reuse of network features.Secondly,attention mechanisms are employed to enhance the effectiveness of features,utilizing high-order statistics to measure and adjust the characteristics of feature maps.Lastly,LCRCA splits convolution layers to create deeper residual modules with the same number of parameters.Additionally,the learning process is transformed from one-stage to multi-stage residual learning,thereby enhancing the feature extraction capability of the residual modules.By implementing these three measures,LCRCA achieves superior spatial resolution enhancement with a lightweight network scale.(4)The challenge of joint bit and spatial resolution enhancement lies in effectively suppressing artifacts caused by both types of resolution degradations while considering their mutual interference.To address this,the dissertation proposes a divide-and-conquer and adaptive fusion-based algorithm for joint bit and spatial resolution enhancement.Artifacts resulting from bit and spatial resolution degradations exhibit noticeable differences,and their blending degrees in different image regions significantly vary.But these degradation processes can be represented as numerical mappings,thus showcasing both homogeneity and heterogeneity in their characteristics.Unfortunately,existing algorithms fail to consider these properties simultaneously and neglect the mutual interference of degradation artifacts,resulting in unsatisfactory reconstruction quality.To overcome these limitations,the dissertation introduces a divide-and-conquer strategy to mitigate the mismatch between the "shared weights" characteristic of convolution and the diverse compositions of artifacts.Additionally,adaptive fusion enables joint optimization of the two resolution enhancements.By leveraging the divide-and-conquer strategy and adaptive fusion,the proposed scheme in this dissertation achieves superior joint enhancement of bit and spatial resolution. |
PDF Full Text Request