Joint estimation image defogging method based on convolutional neural network

Abstract

The invention discloses a single image defogging method, which mainly solves the problems of constrained nonlinear fitting ability, cumbersome training and single applicable scene in the prior art. The scheme comprises: constructing a convolutional neural network consisting of a feature sharing part, a haze atmospheric light value estimation branch and a transmittance estimation branch under the Caffe framework; acquiring a group of fog-free image set J, artificially fogging J to obtain a fog image set I; averagely dividing I and J into multiple paired image groups by batch size, and sequentially inputting in order for 200000 times to the neural network for training; inputting the image I that needs to be defogged into the trained neural network, and outputting the atmosphere light value Aand transmittance T; and calculating according to the atmosphere light value A and transmittance T to obtain a fog-free image J. The invention may well maintain the contrast and color saturation of the restored image, with the peak signal to noise ratio and the structural similarity both superior to the prior art, and may be used for the clear processing of the foggy image.

Description

technical field [0001] The invention belongs to the technical field of image processing, and in particular relates to a method for defogging a single image, which can be used for clearing processing of a single foggy image captured by an imaging system. Background technique [0002] Affected by severe weather such as smog and haze in real conditions, the image quality captured by imaging equipment is generally low, often with low contrast, color shift, and low information recognizability. These degraded images not only affect the subjective experience of the human eye, but also seriously affect the performance of various intelligent visual information processing systems. Therefore, the clear processing of foggy images has very important practical application value. [0003] At present, the image and video dehazing algorithm based on the atmospheric scattering model is a research hotspot, and the key issue is how to estimate the atmospheric light and transmittance. [0004]...

AI Technical Summary

Problems solved by technology

[0010] The purpose of the present invention is to propose a joint estimation image defogging method based on convolutional neural network to solve the problems of restricted nonlinear fitting ability and tedious training in the prior art and too single applicable scene, and fully utilize convolution Neural Network Dehazes Images

Images