Single Image Dehazing Method Based on Mean-Mean Square Deviation Dark Channel under Superpixel Framework

Abstract

Provide a single image dehazing method based on the mean-mean square error dark channel under the superpixel framework. The steps are as follows: Calculate the minimum value matrix I of each color channel of image I according to formula (6) dark , hereafter referred to as the gray matrix I dark ; Obtain the appropriate parameters T and k according to step 1, and calculate the fog removal rate ω according to formula (5); estimate the atmospheric light value A; calculate the matrix img = I / A; perform super-pixel segmentation on the matrix img to obtain the imaged The scene depth d(x) and the scattering coefficient β of the medium in the atmosphere are constant at several Ω i ; for each Ω i Calculate the dark channel according to formula (4) to obtain each Ω i The resulting stitches are stitched together to obtain the dark channel J of the entire image. dark ;By t=1‑ω*J dark Calculate the rough transmittance t; refine the rough transmittance t to obtain the refined transmittance t * ;From the formula J=(I‑A) / t * +A gets the final restored image J. The method of the present invention can keep the fog concentration and depth of field unchanged in the local area represented by the super pixel, overcome the halo effect at sudden changes in the depth of field, and effectively alleviate the color cast problem caused by the infinite depth of field.

Description

technical field [0001] The invention relates to image processing technology, in particular to a method for defogging a single image based on a mean-mean-square-difference dark channel under a superpixel framework. Background technique [0002] Haze, as a common natural phenomenon, is an important cause of image degradation. The mechanism is that the small droplets and aerosols in the atmospheric haze components scatter visible light, making it impossible to pass through the air medium normally, resulting in the deterioration of the imaging effect. Imaging under the influence of smog brings great difficulties to subsequent processing such as segmentation, detection, and recognition. Therefore, how to visually eliminate smog interference and improve image visibility is very necessary. [0003] Existing technologies can be divided into single image and multiple images in terms of data sources. The advantage of multiple images is that they are rich in information, but the diffic...

AI Technical Summary

Problems solved by technology

[0013] (1) How to use formula (2) when solving formula (1), there is a prerequisite, which requires t(x) to be kept constant within Ω(x), according to the previous analysis, t(x) has the same relationship with β and d(x) relationship, and Ω(x) is a square area centered on pixel x, if the selected Ω(x) contains multiple depths of field, this precondition is not satisfied, and Ω(x) will overlap each other when selected, resulting in Inaccurate estimate of t(x)

[0014] (2) When Ω(x) contains multiple depths of field, a halo effect (Halo Effect) will be generated at the sudden change in the depth of field, which is the close-up C f For prospect C b Caused by the occlusion of the C f and C b Caused by the wrong classification of pixel categories, resulting in the failure of defogging in local areas of the image

[0015] (3) In the atmospheric scattering model, J(x)t(x) is called "direct attenuation item", A(1-t(x)) is called "atmospheric light item", by t(x)=e -βd(x) It can be seen that when β is constant, the contribution ratio of these two items to the imaging result shows a trade-off relationship with the increase of d(x), and the dehazing effect of the dark channel priority theory depends on the J(x)t(x) in the whole The proportion of imaging, when J(x)t(x) is dominant or comparable to A(1-t(x)), the effect of defogging is obvious, otherwise, satisfactory defogging effect cannot be achieved

[0021] (1) The assumptions for the application of the median dark channel are not necessarily true, that is, the transmittance cannot be guaranteed to remain constant within Ω(x)

[0022] (2) If there is no depth of field in which pixels are absolutely dominant in Ω(x), the median value cannot effectively suppress the occurrence of the halo effect

[0023] (3) The median dark channel essentially avoids taking the minimum value. When d(x)→∞, the estimated transmittance value is larger, which can suppress the color cast problem in large bright areas to a certain extent, but the effect is still not ideal

Images