Research On Degraded Image Restoration In Medium

Outdoor images are often degraded by scattering medium,resulting in limited visibility,low contrast and saturation,details obscuration and shifted luminance.The clarity and color fidelity of degraded images decline seriously and the information is greatly reduced,which affects the subsequent computer vision applications such as video surveillance,target tracking and object recognition.Most existing image restoration methods use optical model to describe the degradation and focus in proposing different kind of priors to estimate transmission map in the optical model.The recovered result is obtained indirectly from the optical model after estimating transmission map,which ignores the characteristic of the recovered image itself.The result of this strategy is significantly boosting the invisible image noise and color aliasing existing in the input images,which are visually intruding in the recovered results.Therefore,this paper concentrates on the degradation in imaging process to optimize the restoration algorithm and discusses the expression of image characteristic.The main research items are as follow:(1)For suppression of the boosted noise in restoration,a new haze removal method based on regularization is proposed.First,the transmission map is refined by using guided filter with V channel of HSV color space to reduce unexpected detail edges.A new Weighted Total Variation regularization of transmission map is proposed,which considers the spatial information and contributes to improve the accuracy fine details in the haze removal result.To restrain the noise boosted during restoration,Relative Total Variance is introduced as regularization term and an adaptive regularization parameter is designed to adjust the strength of regularization constraint according the depth of region.Experiment results demonstrate that the proposed method can recover the unobservable details in the haze image and suppress the boosted noise during restoration,which effectively improve the overall quality of image.(2)For image haze removal,a new hybrid convolutional sparse representation model is proposed.According to the different characteristics of transmission map and scene radiance,different forms of sparse representation prior are adopted to constrain sparsity.Considering the characteristic of piece-wise smooth,the analysis dictionary is used to represent transmission map.And the synthesis convolutional dictionary is used to represent fine textures of scene radiance according to the characteristic of non-local self similarity.The complementary characteristic of analysis sparse representation and synthesis sparse representation makes the proposed hybrid model be able to effectively recover the textures of scene radiance and suppress the noise by the sparsity constraint.Experiment results demonstrate the effectiveness of the proposed method on fine structures recovery and noise suppression.(3)For underwater image restoration,a new imaging model combined with the Retinex theory is proposed.In the proposed model,the forward and backward scattering effects are modeled in the form of multiplicative attenuation.And the lighting component is separated from the scene reflection to reduce the degradation from inhomogeneous light.In the corresponding solution procedure,the backscattering rate and ambient light are estimated first.And then,an optimization objective function about the attenuated lighting component is established to minimize the influence of external lighting on the obtained scene reflection component,which cooperates the backward scattering information and introduces Total Variation as a regularization term.Experimental results show that the proposed model and method are extensively effective on the degraded image of natural light source and artificial light source.