A Convolutional Neural Network Demosaicing Method Based on Residual Interpolation

Abstract

The invention belongs to the technical field of image processing, in particular to a convolutional neural network demosaicing algorithm based on residual interpolation. In order to solve the problem that the current image restoration method is not accurate enough in the hypotenuse area and the overall resolution of the image is low, this method first uses the edge detection algorithm to restore the green channel, and then uses the residual interpolation method to restore the red and blue planes. After obtaining the residual interpolation results, Finally, the convolutional neural network is used as a correction item to further improve the resulting image. The demosaic result is used as the input of the convolutional neural network, and the residual between the corresponding full-color image and the demosaic result is used as the label of the convolutional neural network. After training Gradually modify the weights of a convolutional neural network. This method not only detects edges in the horizontal and vertical directions, but also increases the detection of edges in the oblique direction, so that the image can be restored more accurately in the oblique edge area, the resolution of the image is improved, and the edge of the demosaic image obtained is clearer.

Description

technical field [0001] The invention belongs to the technical field of image processing, in particular to a convolutional neural network demosaicing method based on residual interpolation. Background technique [0002] Single-sensor color imaging technology is widely used in the digital camera industry. In a single-sensor camera, a color filter array (CFA: Color Filter Array) is covered on the surface of the sensor, and each pixel only samples red and green. One of the three color components, blue and blue. If the full-color image is restored, the two lost color components need to be estimated. We call the process of estimating the lost color as demosaicing. The most widely used color image today is the image of the Bayer CFA model. In the Bayer CFA model, the green pixels are sampled according to the quincunx grid, and the red and blue pixels are sampled according to the rectangular grid. Among them, the green samples are red or blue double the sample size. So far, resear...

AI Technical Summary

Problems solved by technology

However, these methods are equivalent to isotropic low-pass filtering, which will produce significant defects such as edge blur and zipper effect.

In response to this defect, an adaptive direction interpolation algorithm can be used to detect the direction of the horizontal and vertical edges using the gradient, and interpolate along the horizontal and vertical edge directions respectively, but this method will reduce the overall resolution of the image; The method based on the smooth transition criterion of hue interpolates according to the constant assumption of color difference (red-green, blue-green), but the obtained full-color image has a large interpolation error where the green component value changes abruptly.

The residual interpolation algorithm is used to improve on the basis of chromatic aberration interpolation. This method performs interpolation on the residual domain. Compared with the chromatic aberration domain, the residual domain is smoother and the interpolation accuracy is higher, but this method is not accurate enough in the recovery of the hypotenuse area.

Images