Multi-focus Image Fusion Based On Multi-scale Transform

Image fusion is a process of combining two or more images obtained from differentsensors or of different imaging modalities to one image. The fused image reflects themultifaceted information of the source images and is more suitable for human visualperception or computer processing. Therefore, the image fusion technique can be employedto improve the efficiency of using image information and hence the reliability of the targetdetection system can be improved. Currently, multi-focus image fusion technology iswidely used in machine vision, digital cameras, target recognition, medical imaging and soon.In view of the problems of the existed traditional multi-focus image fusion techniquesand upon considering the characteristics of multi-focus images, two types of image fusionalgorithms are proposed in this thesis based on the theory of multi-scale transform.First, an image fusion algorithm is proposed based on the shearlet transform and theadaptive pulse coupled neural networks (PCNN). This algorithm decomposes two alignedimages to be fused with the shearlet transform. According to the difference in the low andhigh frequency coefficients, different fusion rules are used respectively for them. For thelow frequency sub-band, coefficients are fused based on the regional energy whilecoefficients in high frequency sub-band are fused based on the adaptive PCNN fusion rule.Fused coefficients in both low and high frequency sub-bands are then used as inputs to theinverse shearlet transform to obtain the final fused image. Different from traditional fusionalgorithms based on the PCNN where the linking strength is determined empirically, thegradient energy is used as the linking strength of PCNN. In this manner, the linkingstrength can be chosen adaptively according to the gradient energy. In the proposedalgorithm, the high frequency sub-band coefficients are determined according to the timesof firing in the PCNN. Experimental results show that the proposed algorithm of imagefusion achieves better fusion quality than traditional approaches. Second, a new image fusion method is proposed based on the shearlet transform andimage regional features. In this method, to make the best use of the regional characteristicsof the low frequency coefficients, the low frequency sub-band coefficients from differentimages are fused according to a rule based on regional energy. As for the high frequencysub-band coefficients, they are fused according to a novel rule upon comprehensiveconsideration of multiple regional features，such as regional variance, regional averagegradient, regional spatial frequency and so on. Experimental results show that the proposedalgorithm for image fusion outperforms traditional ones.