Research On Subjective And Objective Evaluation Models Of Image Quality Based On Digital Vedio Surveillance Devices

Digital video surveillance devices are special in their video image processing.When the video processing tasks are running, media softwares would employ aseries of image processing algorithms to achieve the effects of noise-reduction, up-scaling or down-scaling, encoding, deinterlacing, etc. Different algorithm parameterconfigurations result in different visible contents of images, literarily differentimage quality. Evaluation of image quality could provide improvement proposelsand directions for image processing algorithm developers, as well as objectiveprinceples to compare image quality of different images, thus the image qualitymonitoring function could be automized. This automation is also a necessary step toself-adaptively ajust image processing procedures, for example, the ajustment ofencoding bitrate and bitrate controling parameters. In order to assess their imageprocessing performance, we proposed a package of subjective and objective imagequality assessment (IQA) method and model, including subjective IQA andobjective IQA MATLAB model. Subjective assessing methods are focused on imagequality assessment fundamental theories, assessing enviroment construction andassessing procedures, while objective assessing methods are focused on thealgorithm model of image quality assessment based on computation, to emulatesubjective assessing effect. In this project, a new method is proposed as SWVIFalgorithm, which is based on popular visual information fidelity assessment methodand improved from the latter. To be differentiated from previous studies on IQAmethods, this project is specified to focus on those unique image artifacts broughtabout in digital video surveillance devices image processing. A typical form ofdigital surveillance device is employed as reference in this project, the Digital VideoRecorder(DVR). For DVR video processing, there exist the following five types ofimage artifacts:(1) the Mosaic artifact resulted from video encoding with loss;(2)the bluring artifact resulted from the noise reduction image processing procedures;(3) the white-noise artifact due to white-noise contamination in the transmission ofsignals;(4) the luminace mismatch artifact;(5) the zig-zag artifact on brims ofimage contents resulted from image down-scaling. To weight different artifacts withdifferent importance, a Scale-Weighted Visual-Information Fidelity (SWVIF) stillimage assessment model emulating the scale-weighing effect of HVS is designed,which is based on widely accepted Visual-Information Fidelity (VIF) model, whichis a method based on steerable pyramid wavelet decomposition and construct thereference image model and distorted image model to grasp the characteristics ofimages. However, the VIF method does not take into account the multi-scaleweighing funtionality of HVS, thus neglecting some weighting preference on different scales during the assessment. The new model has been constructed inMatlab development enviroment and verified on LIVE image database with3different types of image artifact including JPEG, Gaussian Blur and White-Noiseand3evaluation metrices including PLCC, RMSE and MAE are used to assess theaccuracy of the SWVIF algorithm model. It proves to be better in overallperformance than VIF and Multi-Scale Structural Similarity Metric (MS-SSIM),both of which are widely accepted as effective objective IQA models.