Intrinsic Decomposition Algorithm Of Single Image Under Complex Lighting Conditions

The intrinsic image decomposition algorithm aims to decompose the input image into a reflectance and an illumination.The reflectance represents the material information of the object,and the illumination represents the illumination information of each position in the image.In the decomposition process,since there is only one known quantity,but there are two unknown quantities that need to be solved,the intrinsic decomposition algorithm is an under-constrained problem,and more constraints need to be introduced to help the decomposition.Intrinsic image decomposition has important application value in the field of image processing.In order to ensure the authenticity of image synthesis,the intrinsic decomposition algorithm decomposes the reflectance and the illumination.First insert an object in the reflectance,and then insert the inserted Multiplying the reflectance and the illumination pixel by pixel can ensure that the lighting of the inserted object is consistent with the lighting of the scene,making it more realistic in terms of visual effects.In recent years,the research results of intrinsic image decomposition are very rich,but most of the research focuses on scenes with uniform illumination.When the lighting conditions are more complicated,the decomposition algorithms do not behave very well.Because the existing intrinsic decomposition algorithm does not consider the particularity of shadow and highlight areas,the material,texture,and lighting information of some shadow areas cannot be correctly decomposed,and it may also lead to the loss of material and texture information in highlight areas.To solve the above problems,this thesis proposes an improved intrinsic-image decomposition algorithm under complex lighting conditions on the basis of existing algorithms.After experimental verification,the proposed method has good performance in visual effects and quantitative results.The main research contents of this thesis are as follows:1.In order to prevent the shadow from being treated as a material problem,this article first uses the shadow detection method to identify the shadow in the scene,and then uses the shadow mask generated by the shadow detection to correct the original image using the Color Line-based method.The improved image is clustered,and the pixels of the same category have the same material.According to the weight information of different materials,the weighted least squares optimization algorithm is used for the original image to perform the initial illumination estimation,and the result of the illumination estimation is used as the shadow constraint.2.In order to solve the problem of image information loss around highlights in the scene,this thesis uses the random forest to detect the highlight areas.Because the model generated by directly inputting the labeled images and labels into the random forest takes up memory,this thesis uses superpixels to extract features from the input images,which reduces the model size and speeds up training and training while maintaining good prediction effects.The speed of prediction.3.Existing intrinsic image decomposition algorithms tend to smooth and sparse the reflectance layer,resulting in the texture information of the illumination layer being classified into the illumination layer.In order to retain the texture information of the reflectance layer,this thesis saves the texture of the reflectance layer in the process of iterative solution based on the gradient information,so that the color and texture of the reflectance layer is more visually pleasant.In order to verify the effectiveness of the algorithm in this thesis,this thesis has carried out visual effect evaluation and quantitative evaluation on two public data sets and algorithms that have performed well in recent years.The reflectance layer and the illumination layer are evaluated using the IIW and SAW data sets respectively.From the perspective of visual effects and quantification,the experimental result not only obtains satisfactory results in evenly illuminated scenes,but also performs well in complex illuminated scenes.From the visual point of view,the decomposition result of this thesis is better than many excellent intrinsic image decomposition algorithms in recent years.Finally,this thesis also applies the intrinsic-decomposition algorithm to image synthesis and color-editing,verifying that the algorithm in this thesis can ensure that the inserted and color-edited objects are consistent with the lighting condition in the scene.