| Most digital cameras and other color imaging devices are structured with a single sensor surface covered with different color filters to save space and cost.As it captures only one color per pixel position,the color information that was unsampled must be reconstructed by the image demosaicing technique.That is,the image demosaicing technique is the process of reconstructing RGB information from a sparsely sampled mosaic image.The image quality of color imaging devices represented by digital cameras mainly depends on the reconstruction precision of image demosaicing algorithms.In addition,image demosaicing tasks are the basis for a variety of downstream vision tasks,and efficient and accurate reconstruction results can greatly improve the performance of downstream vision tasks.With the development of deep learning,many highperformance demosaicing algorithms have been proposed in the field of image demosaicing,but there are still problems such as difficult trade-off between high quality and low cost,and poor generalization.This study considers that image demosaicing has the characteristic of high degradation,such as the location of the pixel information retained in the sparsely sampled mosaic image is discontinuous.If only a simple convolution operation is used to consider the neighborhood location characteristics,it is difficult to fully extract global features,thereby affecting the reconstruction efficiency of the network model.Given that the non-local operation considers the correlation between each pixel information and can increase the perceptual field to extract global features,for different application scenarios and the problems of existing demosaicing algorithms,this paper proposes the following two algorithms based on the non-local operation:(1)In this paper,we propose a demosaicing network model based on lightweight non-local block and dynamic weight learning to achieve high reconstruction performance with faster reconstruction speed for deployment in small mobile devices such as mobile phones and surveillance.Firstly,this algorithm proposes a mix-attention block(MAB)that extracts both global and local features,which facilitates the reconstruction performance by reducing imaging artifacts.Furthermore,in order to enhance the network performance without increasing the network parameters,this algorithm adopts a dynamic learning strategy to optimize the weights of each feature according to the position of each attention branch in the network.Finally,extensive experiments demonstrate that our proposed network model is better able to reconstruct colors and details close to those of real images,and better balance the reconstruction accuracy and running time.(2)The proposed demosaicing network model based on region-level non-local block and residual aggregation is more suitable for scenarios with higher accuracy requirements compared to Scheme 1.Firstly,this algorithm proposes a region-level non-local operation module,which divides the features extracted by front-end convolution into four subregions of 2 × 2.Then the sub-regions are subjected to non-local operations separately.The feature correlation at each location is computed only in the pixel neighboring regions to improve the representation capability of the network.Second,this algorithm uses residual aggregation to make full use of the front-end purer residual features.The different spatial features extracted from different residual blocks are completely integrated at the end of the residual blocks to improve the network reconstruction efficiency.Extensive experimental results and parameter analysis show that our proposed network model achieves better performance.Finally,in order to demonstrate the universality of our proposed network model for different types of image reconstruction,the network model is applied to the field of RGB-NIR image demosaicing for performance evaluation,and it is demonstrated that our proposed network model can be applied to different types of image reconstruction with better generalization ability. |
PDF Full Text Request