Research On Demosaicing Algorithm For Bayer Image

In consideration of design cost and economic cost,most modern color digital cameras are equipped with a single sensor to acquire images.The Bayer image obtained by single sensor has only one of three colors of red,green and blue at each pixel.To get a full color image,it is necessary to estimate the missing color component value at each pixel.The above process is usually called Bayer image demosaicing.The result of the demosaicing algorithm is directly related to the quality of the color image obtained by the color digital camera,and it is also a prerequisite for advanced image processing technologies such as object detection and image recognition.Therefore,the study of efficient demosaicing algorithm still has important practical value.In view of the shortcomings of the current demosaicing algorithm in terms of time and accuracy,two demosaicing algorithms are proposed in this paper.1.A new high-quality and low-cost demosaicing algorithm based on gradient weight fusion is proposed for time critical scenes.With the improvement of computing power of central processing units(CPUs)and graphics processing units(GPUs)in recent years,the image quality generated by the classic low-cost demosaicing algorithm,which is widely used,has gradually failed to meet people's visual needs.As a classic high-quality linear interpolation algorithm,Malvar's High Quality Linear Interpolation(HQLI)algorithm is the most widely used in practical applications.It is widely used in time-critical scenes or as an input to the initialization of complex demosaicing algorithms,but the algorithm is prone to artifacts at the edges and the visual effect is not satisfactory.After studying the advantages and disadvantages of the HQLI algorithm,we designed a new high-quality and low-cost demosaicing algorithm that fuses edge gradient weights.In this algorithm,an improved block gradient method is adopted,and the gradient is non-linearly mapped to the direction weight through the logic function,so as to obtain more accurate weight.In the calculation of direction candidate interpolation,a larger neighborhood range is used,so a more accurate candidate interpolation is obtained.Finally,the direction weights and direction candidate interpolation are adaptively fused to reconstruct a high-quality demosaicing image.And compared with a lot of demosaicing algorithms that need to interpolate the green channel first and then interpolate the red and blue channels,our algorithm can obtain all the missing color components in the Bayer image in one interpolation process.Tested on the standard datasets Kodak and McM,the accuracy of our algorithm has greatly improved compared to the HQLI algorithm.Compared with the recently proposed highquality and low-cost algorithm LED,our algorithm obtains higher PSNR and image quality with less time-consuming.2.In the scenario where accuracy is priority,a quality improvement algorithm of demosaiced image based on deep residual network is proposed.Although the high-quality and low-cost method we proposed completes the demosaicing task with high efficiency and superior performance,it still has artifacts that affect vision.In application scenarios where accuracy is more important than speed,higher quality images are still our primary goal.Therefore,we use the results of the proposed high-quality and low-cost algorithm as the input of the convolutional neural network,and correct the color values and reduce artifacts by training the convolutional neural network,so as to improve the quality of the demosaiced image.We propose a new twostage improvement algorithm for the quality of reconstructed images based on deep residual networks.The goal of the first stage is to use the deep residual network to improve the quality of the green channel.The second stage uses the obtained green channel to guide the improvement of the quality of the red and blue color channels.Inspired by the Color Difference Recovery CNN model recently proposed by Niu and Ouyang,we transfer the reconstruction of the red and blue channels to the green-red and green-blue color difference planes,and optimize the color difference plane with a deep residual network to reconstruct high-quality color image.Finally,confirmatory experiments were carried out on the standard datasets Kodak and McM.The results of subjective and objective comparison with the current several algorithms show that the image reconstructed by the algorithm in this chapter has reduced artifacts and greatly improved the quality.