Research Of Image Fusion Based On Intelligence Optimization And Visual Saliency

Image fusion has become a hot research field with the development of information, sensor and image processing technology. Image fusion integrates multiple images into a single image, which describe the same scene. These multiple images are obtained by different sensors or one sensor in different ways. The fused image has more useful information and decribes the scene more comprehensively and accurately. Therefore, image fusion has been widely applied to many fields such as machine vision, target identification, medical diagnosis, remote sensing and so on.This thesis analyzes the development and problem of image fusion in spatial and frequency domain. Intelligent optimization, multi-scale transform and visual saliency detection are studied and applied into multi-focus or infrared and visible image fusion. The main contents and innovations are summarized as follows:(1) A sharpness estimation method based on total variation and modified sum modified laplacian is proposed in spatial domain. Combining the global feature of total variation with the detail feature of modified sum modified laplacian, the proposed method accurately estimates the sharpness area of multi-focus images with low computational complexity for image fusion processing.(2) A multi-focus image fusion method based on artificial fish-swarm optimization is proposed. The sharpness estimation method is used to select approporiate image blocks to reconstruct the initial fused image. According to the fitness function, artificial fish-swarm optimization algorithm uses its global feature to iteratively search the optimal image block size to reconstruct final fused image.The performance of the fused image has been enhanced and the block artifact has been suppressed.(3) A multi-focus image fusion method based on sharpness estimation is proposed in frequency domain. The source images are decomposed by double density dual-tree complex wavelet transform and the sharpness areas are obtained by local energy contrast of the transform coefficients. Then the filter technology further optimizes the sharpness area. The fused image is constituted with the selected sharpness areas combing the image structure similar characteristics.(4) A multi-focus image fusion method based on dicrete cosine tranform is proposed. For real-time application, the spectrum entropy ratio of dicrete cosin transform is used to select focus blocks in order to constitute the fused image. The propsed method can effectively eliminate boundary blur and suppress block artifact with low computational complexity.(5) Combined with multi-scale and multi-objective optimization, a multi-focus image fusion method is proposed. Firstly, the source images are decomposed by non-subsampled shearlet transform. Then the multi-objective biogeography based optimization is used to find the best possible optimal solution of the fused image for balancing multiple objective evalution indexs. The performance of the proposed method is superior to the traditional methods based on multi-scale transform which has more sharp edge and texture.(6) A saliency detection of infrared image based on non-subsampled shearlet transform is proposed. Because of localization, scale and orientation sensitivity of non-subsampled shearlet transform, the proposed method has better detection results than the previous methods in frequency domain and effectively highlights the targets in the infrared images. Then an infrared and visible image fusion method based on the saliency detection is proposed. Different fusion rules for salient and non-salient region are able to enhance the contrast of fused image and retain the scene and texture detail of visible image and the infrared target has integrity and clarity.