Image super-resolution reconstruction method of supervised convolutional neural network based on multi-scale feature extraction fusion

Abstract

The invention discloses an image super-resolution reconstruction method of a supervised convolutional neural network based on multi-scale feature extraction fusion. The method comprises the followingsteps: image preprocessing, image feature extraction and image reconstruction; in the image feature extraction step, a plurality of MSB modules are adopted, in the MSB modules, feature extraction is carried out on an image by adopting convolution layers containing convolution kernels of different sizes, feature repeated learning is carried out in a dense connection mode, and a supervision layer error function is designed in the model and used for assisting and correcting reconstruction errors of the model. According to the method, the extracted feature map is processed on different scales, sothe adaptability of the model is enhanced; multi-channel propagation of information is achieved, the convergence speed is increased, and the gradient disappearance phenomenon is relieved; and an auxiliary supervision error function is added, so the back propagation of the gradient is enhanced, extra regularization is provided, the problem of gradient disappearance in a traditional algorithm is effectively solved, and the precision of the algorithm is improved.

Description

technical field [0001] The invention belongs to the technical field of image processing, in particular to an image super-resolution reconstruction method based on a multi-scale feature extraction and fusion supervised convolutional neural network. Background technique [0002] With the rapid development of digital multimedia technology, digital images have become an important form of information transmission. The level of image resolution and the amount of useful information carried are directly related to the depth of people's cognition of information, which makes people have higher and higher requirements for image and video quality. However, in the actual process of acquiring digital images, it is often affected by many factors, resulting in the degradation of the acquired image quality, such as various noises inside and outside the imaging system, and in some special occasions, the data needs to be transmitted, saved, and down-sampled. Or compressed, the resolution and ...

AI Technical Summary

Problems solved by technology

But it also has some defects: First, the SRCNN algorithm only uses three convolutional layers, and fewer convolutional layers lead to certain limitations in the reconstruction performance of the trained model

Second, the receptive field of the SRCNN algorithm model is small, and the local receptive fields between different convolutional layers can constitute the context information association of neurons in adjacent layers. The SRCNN algorithm only uses three convolutional layers, resulting in too large receptive fields of the model Small (13×13), the model cannot make full use of the information between different convolutional layers, which affects the reconstruction performance of the model

Third, the SRCNN algorithm has poor adaptability in extracting features, and SRCNN does not integrate features, resulting in poor model reconstruction performance

In addition, in recent years, a large number of algorithms have emerged to improve the SRCNN algorithm, but they all have certain shortcomings, such as simply relying on increasing the number of network layers to improve the quality of image reconstruction, because in the network training process, The parameters in the model are continuously updated iteratively through gradient backpropagation, so that the feature extraction tends to reduce the error. When training a model with more convolutional layers, the weight parameters of the first convolutional layer may not be obtained. The results of the gradient backpropagation are effectively updated, resulting in the quality of the image reconstructed by the model is not as good as the model with fewer convolutional layers. This phenomenon is also known as the gradient disappearance phenomenon.

Images