Non-uniform noise removal method based on stepped multi-scale convolutional neural network

Abstract

The invention discloses a non-uniform noise removal method based on a stepped multi-scale convolutional neural network, and the method comprises the steps: firstly constructing a noise reduction convolutional neural network model, and carrying out the training of the noise reduction convolutional neural network model based on an existing noise image and a clear image; the noise reduction convolutional neural network model comprises a cavity convolution combination processing module, a stepped multi-scale noise identification module, a multi-scale feature fusion module and a deep learning noise reduction module which are connected in sequence, and the output of the convolution combination processing module is connected with the input of a mixed cavity convolution combination processing module of the deep learning noise reduction module; and finally carrying out noise reduction processing on the noise image based on the noise reduction convolutional neural network model obtained by training, and outputting a noise-reduced image. According to the scheme, a stepped multi-scale network structure is designed based on parameter lightweight thinking, 1 * 1 convolution and 3 * 3 cavity convolution are introduced, image multi-scale information can be fully acquired on the premise that parameter quantity is not increased sharply, and more image detail information is reserved while non-uniform noise is removed.

Description

technical field [0001] The invention belongs to the field of image data processing, and in particular relates to a non-uniform noise removal method based on a stepped multi-scale convolutional neural network, which can be applied to point, strip, and plane non-uniform noise image noise reduction. Background technique [0002] In conventional image acquisition systems, affected by hardware and the environment, non-uniform noises of various shapes and scales are likely to exist in images, such as remote sensing images containing stripe noise, underwater video images containing impurities, etc. The subsequent development of image segmentation and target recognition will have adverse effects. [0003] There are mainly three kinds of traditional image noise reduction methods: methods based on grayscale information statistics, such as histogram matching method and moment matching method. The problem of “unclean”; methods based on digital filtering mainly include spatial domain fi...

AI Technical Summary

Problems solved by technology

The classic neural network noise reduction convolutional neural network model, especially the Unet model, keeps the network input and output tensor sizes consistent through the process of pooling and deconvolution. Among them, pooling is to reduce the calculation by compressing feature data. However, it is inevitable to lose image detail feature information when compressing features, so when dealing with non-uniform noise, there will be a problem of blurred image details

In addition, as the application of convolutional neural network gradually deepens, its structure gradually evolves from a single, multi-layer structure to a complex, deep structure, and the amount of parameters is increasing day by day, while the training of parameters is limited by the size of sample data, over-parameterization will Lead to model overfitting and reduce model generalization ability

Images