Retinex algorithm image defogging method based on neural network

The invention relates to the technical field of image processing, in particular to a Retinex algorithm image defogging method based on a neural network, and the method comprises the following steps: S1, data early-stage conversion: converting a pixel value of an original image from an integer domain to a logarithm domain, reducing the operand of a subsequent algorithm, firstly decomposing a color image into three grey-scale images of R, G and B, and then converting the three grey-scale images, namely s (x, y) = log (1 + S (x, y)); s2, initializing a matrix which has the same size as the original image S (x, y) and has a constant t, and t is the mean value of the brightness of the original image; s3, the value of S = 2P is solved, P = fix [log2min (m, n)-1], and S is the maximum distance between the target point and the two comparison points; s4, calculating pixel points on the path; s5, making S =-S/2; the distance between the two comparison points in each step and the target point is shortened to be half of