Model Robustness Analysis And Quality Assessment Model Research Of Depth Images

At present,depth images have received increasing attention in the computer field,and there are multiple channels for obtaining depth images.As an important component of 3D reconstruction technology,depth images have begun to gradually move into people’s lives.Unlike natural images,depth images are also known as distance images,which reflect the distance from the scene within the image to the camera.Natural images do not have accurate distance data and can only rely on human visual judgment.Generally,high-quality depth images are relatively difficult to obtain,especially when there are problems such as exposure during the shooting process,and the loss of details encountered during image transmission.Images that are inherently not of high quality become increasingly blurred.This problem has established the rise of the field of image super-resolution and image enhancement.Based on depth images,this article will fully study depth images from two aspects: super-resolution and quality evaluation.The specific research work is as follows.(1)Depth image super-resolution(DISR)is becoming more and more important in the field of computer vision.Although great progress has been made in this research topic,the robustness of DISR models is not sufficiently investigated,which is of great importance in the real applications.Accordingly,in this thesis,we make an initial attempt to investigate the robustness of DISR models.Specifically,we test their generalization ability when the input depth image suffers from visual quality degradation.To facilitate this study,we construct a large-scale depth image dataset in which the reference depth images are perturbed to generate the degraded depth images automatically.Then,we test six top-performing DISR models on the constructed dataset and then compare their strengths and weaknesses.By conducting comprehensive experiments,we find that depth image super-resolution models perform poorly on Gaussian noise,and that the higher the level,the lower the quality of the predicted depth map.Furthermore,some DISR models only outperform at lower magnifications(such as 2×and 4×).(2)For the depth image data set we built,we propose an image quality assessment model to predict the quality of the depth image.According to the same distortion type and different distortion levels of the same reference image,we use paired depth distortion images as input to calculate the prediction accuracy(Accuracy,ACC).And compared with the FR(Full Reference,FR),NR(No Reference,NR)and image quality assessment based on deep learning methods.Through experiments,we found that,Compared with the current traditional FR,NR and depth image quality assessment models methods,this model has very obvious improvement in performance.This is due to the fact that the established depth learning network fully extracts the edge and other features of the depth image,and the corresponding ranking learning training method also plays a crucial role in the performance of the model.The method proposed in this thesis has achieved good prediction results and remarkable performance,which also further proves the complexity of the proposed large depth image database and the effectiveness of the proposed depth image quality evaluation method.