Research On Semantic Image Coding And Image Quality Assessment Methods

Digital images contain lively and abundant visual information and play an impor-tant role in our daily life and work. In recent years, as the arrival of the age of big data that is accompanied by massive amounts of images, the issues of how to efficiently com-press images while minimizing the distortion of the semantic information, and how to effectively assess the visual quality of images, have become more and more important in the field of digital image processing.Image coding has always been one of the most classical research topics in the field of image processing. Traditional image coding methods, such as JPEG and JPEG 2000, focus on the rate-distortion performance. They aim at minimizing the distortion of image quality at a given bitrate, where the image quality is usually measured by Peak Signal to Noise Ratio (PSNR) or Structural Similarity Index Measurement (SSIM). Typically, the image encoder assumes that human being is the information receiver. However, in many application scenarios, such as mobile visual search, event detection, object tracking, the compressed images are analyzed by computer algorithms. In these scenarios, the encoding process should concentrate on preserving the information that is crucially important for image analysis applications, so as to guarantee the performance of visual analysis task.Image quality assessment (IQA) is another important research topic in the field of digital image processing. An effective image quality assessment approach can better guide the image coding method and improve the encoding efficiency. Existing IQA methods mainly exploit low-level features related to image quality. As a consequence, the high-level semantics that has great impact on visual quality evaluation is neglected. To overcome the limitation of traditional image coding methods and the shortcoming of existing image quality assessment methods, we propose a semantic image coding method and a no-reference image quality assessment approach by exploring the seman-tic information of image content. The main contributions and innovations of this dis-sertation can be summarized as follows:(1) This dissertation proposes a saliency-aware semantic image coding method for mobile visual search. By utilizing saliency detection, we obtain the saliency map that represents the importance of different image areas. Then, we use a threshold segmenta-tion method to generate the salient region that contains important semantic information. During the encoding process, we assign more coding bits to salient region based on the region of interest coding approach, so as to preserve the sematic information that is useful for image analysis. Experimental results indicate that, to achieve similar image retrieval accuracy, compared with traditional JPEG 2000 coding, the bitrate saving of our proposed method can achieve 12.25% or 17.55%.(2) By analyzing and utilizing the high-level semantics implied in global image content, this dissertation proposes a no-reference image quality assessment method based on global and local content perception. We utilize the deep convolutional neural network (DCNN) that imitates the human brain to understand the semantic information of global image content. Based on the visual attention and the multi-channel filtering properties of human visual system, we exploit salient region extraction and Gabor filters to extract the characteristic features associated with local image content. By integrating the global and local content perception, we attempt to design the image quality assess-ment method that is accordant with human perception.