A super-resolution reconstruction method based on feature fusion of dual-channel convolution network

Abstract

The invention discloses a super-resolution reconstruction method based on feature fusion of dual-channel convolution network, which comprises the following steps: a dual-channel convolution network isbuilt based on dense convolution network of different convolution cores; The dual channel convolution network comprises: two sub-channels, each sub-channel adopts the structure of dense connection network, the structure is generated by cascading a plurality of dense connection blocks, each dense connection block is composed of a 1* 1 convolution layer, a 3 *3 convolution layer and a hop layer connection, and a PRELU layer is used as a nonlinear activation function in front of each convolution layer; The weighted L1 norm is used as the loss function. After each sub-channel, the image is reconstructed with super-resolution, and the loss function is calculated and the parameters of the model are optimized. The loss function is the weighted sum of the calculated loss function of the sub-channel output image and the calculated loss function of the whole output; Input any size of low-resolution image, load the trained model, and output the reconstructed high-resolution image.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a super-resolution reconstruction method based on dual-channel convolution network feature fusion. Background technique [0002] Image super-resolution reconstruction technology refers to the technology of reconstructing small-scale low-resolution images into large-scale high-resolution images through computer processing. The problem of image super-resolution is a morbid algorithm problem. On the one hand, super-resolution aims to recover more high-frequency features, so that the texture and outline of the reconstructed image are clearer and the details are richer; on the other hand, super-resolution The rate cannot create more pseudo-details at the expense of image accuracy. Since similar but different images may be down-sampled into the same low-resolution image, super-resolution is a pathological problem that cannot converge accurately. In order to improve the recons...

AI Technical Summary

Problems solved by technology

The problem of image super-resolution is a morbid algorithm problem. On the one hand, super-resolution aims to recover more high-frequency features, so that the texture and outline of the reconstructed image are clearer and the details are richer; on the other hand, super-resolution rate cannot create more false detail at the expense of image accuracy

[0004] However, as far as deep learning is concerned, image features can be abstracted into two concepts: local features and global features, while networks such as VGG and ResNet do not fully extract local features of images.

Although this classical network shows excellent performance in classification, detection and other issues, the super-resolution problem requires more sufficient local features to provide a reference for prediction, which has become a constraint for traditional networks in the field of super-resolution

[0005] At the same time, the traditional classic network adopts a single-channel network architecture, and feature fusion is rarely used. Considering that the features extracted by different convolution kernels of the convolutional network are complementary, the lack of a reasonable feature fusion module will lead to a decline in the quality of image reconstruction. and a decrease in robustness

Images