Research On Image Quality Evaluation Algorithm Based On Semantic And Multi-scale Features

Image quality assessment is an important tool to ensure the quality of digital image processing and transmission,and can help users and developers better understand and optimize digital image applications.Real-world images that people capture in their daily lives are often affected by environmental or equipment parameters,resulting in motion blur,exposure,and other distortions,resulting in image quality degradation.Accurately assessing the quality of real distorted images without a reference image is a challenge because of the different content and distortion types of real distorted images and the difficulty in obtaining undistorted images.Also,with the diversification of mobile digital terminals,images need to be retargeted to fit different terminal sizes.The different dimensions of the retargeted image and the original image resulting in twisting or stretching on the retargeted image that produces distortion types that are essentially different from the common distortion types,and therefore,an effective algorithm for evaluating the quality of an image with common distortion types cannot accurately assess the quality of the retargeted image.Therefore,the design of an effective evaluation algorithm for assessing the quality of retargeted images is an important topic in the field of image quality assessment.To address the above problems,the main research contents and innovations of this thesis are as follows:(1)A no-reference image quality assessment algorithm based on semantic and multi-scale features is proposed for different image contents and distortion types in real distorted images.Most of the distortions in real distorted images exist in local areas of the images.In this thesis,a multi-scale local feature extraction network is proposed to extract local features of images,and an image semantic sensing network based on image semantic features is proposed.Also,an attention mechanism is added to the image semantic feature extraction.In this thesis,experiments are conducted on three real distorted image databases,and the experimental results show that the algorithm is effective.(2)For the problems of geometric distortion features and image information loss in retargeted images,a retargeted image quality assessment algorithm based on semantic and multi-scale features is proposed.The types of distortion in retargeted images mainly include geometric distortion,image content loss,and shape distortion.In this thesis,the multi-scale geometric distortion features and image content loss of retargeted images are calculated using a similar transformation matrix and use a convolutional neural network to extract the deep semantic features,which reflect the high-level features of the images.Meanwhile,to solve the existing problem that there are few and small retargeted image databases,a retargeted image database and subjective evaluation system are constructed in this thesis.Experiments are conducted on the public databases and the constructed database,and the experimental results show that the proposed algorithm can effectively evaluate the quality of retargeted images with high accuracy.