Enhancement of underwater images based on neural network

The two frames of the U-net connection convolution network based on a physical model of underwater imaging, each U-net consists of a coding-decoding network, an attention mechanism is added to the first U-net, Is a characteristic information. By processing an input image and a gray image of it, each layer of the decoding structure is output to the second U-net and a transfer image is estimated. And processing of underwater image estimation using the second U-net to compensate for the underwater image after red light estimation, and the connection of the output of the characteristic information by the first U-net in the decoding structure to ensure that the image detail of the recovery process is not lost. Finally, divide the parent imaging model by the transfer diagram to obtain the final result. This method can effectively improve the underwater image with serious color distortion and serious atomization and at the same time preserve the image detail.