Research On Light Field Image Quality Assessment Based On Deep Learning

As a new visual information carrier,Light Field Image(LFI)has been widely concerned in the field of computer vision in recent years.Compared with traditional 2D image,light field image can simultaneously obtain spatial and angular information of the scene,which makes it applicable in many aspects.Such as depth estimation,refocusing,three-dimensional reconstruction,etc.However,in actual use,different types of distortion will inevitably be introduced in multiple image processing stages of light field image.These distortion factors will reduce the visual quality of light field image and affect the user’s perception experience.Therefore,it is very important to evaluate and optimize the quality of light field image,which can help us better understand and improve the processing process of light field image,improve its visual quality and application effect.In this thesis,the main work of Light Field Image Quality Assessment(LFIQA)is as follows:(1)A LFIQA model based on deep learning is proposed.The network is divided into upper and lower layers.The upper layer network extracts the features of the light field sub-aperture image,and the lower layer network extracts the features of the epipolar plane image,so as to analyze the spatial domain and angular domain features of the light field image.The information of the two is fused in the channel dimension,and the quality score is generated after regression model.It has achieved outstanding results in comparative experiments of SROCC and PLCC with other state-of-the-art algorithms such as BELIF,NR-LFQA,Tensor-NLFQ,demonstrating its strong competitiveness.Subsequent ablation experiments have confirmed the effectiveness of feature fusion.(2)A new LFIQA model based on Vision Transformer is proposed.Since the quality loss of light field image is not only reflected in the spatial domain,but also in the angular domain,the proposed network model focuses on analyzing the sub-aperture image of the light field to reflect the spatial quality change,and stacks three subaperture images of the light field to analyze the quality change in the angular domain,and combines the information of the two to reflect the overall mass attenuation.Experimental results on three commonly used light field quality assessment datasets indicate that compared to the latest evaluation algorithms such as DSA and Xiang’s,this network model achieves better results in terms of SROCC and PLCC metrics.