| Under poor weather conditions such as foggy days,the image acquired by outdoor image acquisition equipment have problems such as reduced texture clarity,reduced color contrast and color offset.Since it is difficult for computer vision system to accurately extract effective information from degraded images,which greatly increases the difficulty of subsequent processing of the system,it is of great research significance and practical value to use efficient image defogging technology to restore clear images to the maximum extent without losing image details or introducing additional interference.In this paper,based on the imaging mechanism of foggy images and the main factors that cause visibility reduction,the image defogging method is studied from the physical point of view on the basis of atmospheric scattering model.The details are as follows:1.In order to restore the image tone better,this paper proposes a wavelength dependent image defogging method based on Mie scattering theory on the basis of the difference in the scattering degree of fog drop particles to images in different wavelength ranges.The calculation shows that when the visibility is 0.1-0.2km,referring to thick fog,the mean standard deviation of the extinction coefficient of light in each color channel of advection fog and radiation fog is 0.0015 and 0.0056,respectively.When the visibility increases to 0.5-1.0km,the standard deviation decreases to 0.00069 and 0.0011,respectively.As the degree of difference decreases,the destructive effect of fog on tonal balance of image decreases.Aiming at the influence of ambient light scattering,the partition iterative estimation is adopted,and it is found that the white light noise value is below 0.09(normalized pixel value)under the condition of uniform mist.When the fog increases to medium concentration and presented non-uniform distribution,or the overall amount of ambient light interference increases due to the widening of visual field interval,the noise value can increase to more than 0.15.When the fog is so thick that it is difficult for human eyes to distinguish objects,the noise value reaches about 0.34.It is found that compared with the comparison algorithm,the proposed method significantly improves the image information entropy and average gradient under the condition of no distortion.Especially for the fog image in the real world,the color information entropy is increased by 15.62%on average,and the average gradient value is increased to twice the original image,achieving the balance between the color of the scene and the clarity of the structure.2.In order to solve the halo effect in images with uneven illumination such as strong local ambient light or obvious artificial light source,this paper proposes an image defogging method based on light source segmentation and point diffusion function.On the basis of light source segmentation,the fuzzy effects of fog layer,halo layer and white noise layer in light source area and non-light source area are eliminated respectively.It is found that this method can improve the image average gradient by 9.38%on the basis of solving the attenuation problem through the color separation channel,improve the structural clarity and suppress the color bias of the light source area.For the synthetic foggy images with the original clear images as the control,the peak signal-to-noise ratio(PSNR)and structural similarity index(SSIM)reach the average value of 30.69 and 0.91,respectively,and the stability is better than that of the comparison algorithms.3.Finally,the image defogging method proposed in this paper is applied to road vehicle detection.Combining the YOLO V3 target detection algorithm with the two proposed image defogging methods,the presence detection of vehicles on the image before and after defogging is carried out.The experiments show that for the scenes of single vehicles,no multiple vehicles shielding each other,and multiple vehicles severely shielding each other,the vehicle identification rate is significantly improved after the defogging process based on light source segmentation and point difusion function.In addition,compared with other comparison algorithms,this method,as a preprocessing method for target detection,can better improve the road vehicle identification rate.Under the condition of no serious occlusion and non-severe haze,the vehicle identification rate can reach more than 90%when the shooting distance is relatively short,about 100 meters. |
