Research On Multi-Resolution Image Fusion Method At Pixel Level

The theory of multisensor data fusion (MDF) was proposed in 1970s. Multisensor data fusion refers to the synergistic combination of different sources of sensory information into one representational format. With the development of technology and diversification of channels to obtain information, Explanation and evaluation of amount of information are demanded. In some special situations, data generated from single sensor is unsatisfactory and people realize the importance of multisensor data fusion more and more. MDF acts as an important position in the local war last decade. From then on it is notified widely by the whole world. Many countries have invested much human resources, material resources and financial resources in researching MDF. MDF has become a hot research area.Image fusion is a field of MDF in which research object is image. Image fusion is a comprehensive modern high technology including many subjects, such as sensor technology, image processing, signal processing, computer, artificial intelligence, etc. We use the term of image fusion to denote a process. Such process generates a single image which contains a more accurate description of the scene than any of the individual source images. This fused image should be more useful for human visual or machine perception. The image obtained from many sensors includes redundant information and complementary information. The redundant information can ensure that the fusion system can obtain the accurate information when some sensors cannot normally work. The complementary information is the fundamental condition of image fusion. Image fusion is not simply to superpose images but can produce new images including more information.Application of multisensor data fusion was firstly used in the military field. Because of the wide applicability and practicability of image fusion's thought, it can transform into civilian application field, such as object detection, automatic target recognition, remote sensing, computer vision, smart buildings, robotics, battlefield surveillance, guidance and control of autonomous vehicles, monitoring of complex machinery and meteorological imaging. Image fusion can obtain wide economic benefits and social benefits.Image fusion is generally performed at three different levels of information representation; these are pixel level, feature level and decision level. The pixel level multisensor fusion is main research field at home and abroad. The image fusion algorithm at pixel level has two categories: one is based on spatial domain and the other is based on transform domain. The image fusion algorithm based on transform domain is the hot research area. Image fusion method based on spatial domain means to perform integration directly in gray space. It is a relatively simple method. Image fusion techniques based on spatial domain mainly include Intensity-Hue-Saturation (IHS) transform, weighted fusion, fusion based on statistical and fusion based on neural network. Image fusion method based on transform domain firstly transforms the image pixels before adopting fusion stategy. The transform algorithms include pyramid representation, wavelet transform (WT) and so on. Lately, wavelet theory has gained much popularity because wavelet transform provides directional information while the pyramid representation doesn't introduce any spatial orientation in the decomposition processes.The process of image fusion involves data acquisition, image preprocessing, image registration and image fusion. The literatures about image fusion are mostly describing image algorithms. Image fusion assessment is also important for a whole fusion strategy. It grossly involves subjective fusion evaluation and objective fusion evaluation. The subjective fusion evaluation depends on the personal subjective sensation. But the personal subjective sensation has a bad influence on the assessment result. The objective fusion evaluation can overcome the influence of the personal subjective sensation. The statistical parameters of fusion image are used by the objective fusion evaluation. We can choice difference evaluation criteria according to difference fusion requirement. The commonly used objective evaluation criteria are information entropy, cross entropy, mutual information, mean value, root mean square error (RMSE), average gradient, spatial frequency, structural similarity, peak signal to noise ratio (PSNR), etc. In this paper, we focus on the methods of pixel level multisensor image fusion. The algorithm based on transform domain is emphasized. According to the questions to be resolved in image fusion, several novel algorithms presented in this paper are respectively image fusion method based on 2D discrete wavelet transform, clustering fusion method based on RBF neural network and multi-resolution fusion method based on chaotic neural network.A novel global energy merging scheme is presented according to the question of distortion in fusion. At first, multi-resolution wavelet decomposition on each source image is performed, and then the region energy is calculated. The weighted average decision rule or the choose decision rule chose by global match measure is used to fuse wavelet coefficients. The three high frequency coefficients of the merging image produce through comparing match measure of several source images. And the low frequency coefficient lies on the choice of the three high frequency coefficient. Finally, by applying the inverse wavelet transform the final fused image is obtained. Experiments showed that the proposed wavelet image fusion method can have better performance than conventional image fusion methods.Aiming at improving the clustering fusion performance, a new method is developed to apply RBF neural networks to clustering fusion. The clustering of every original image pixel is obtained by RBF neural networks and nearest neighbor clustering method. For each cluster center, the membership of every fused image pixel is adopted as the weighting coefficient of the weighted strategy, which is used to obtain the fusion image. The membership is obtained by maximum rule. The weighting rule is used in fusion based on membership. The original data set is chosen as the candidate set of nearest neighbor clustering algorithm, and the center set of hidden units are established dynamically. The proposed method is a faster and accurater clustering fusion algorithm. The advantages of RBF network training is reflected in this method including no iteration, few training parameters, high training speed and applicability for clustering. The experiment results show that the proposed method can achieve better performance than self-organizing feature map (SOFM) neural networks method. The influence of the various widths is discussed in this paper. Furthermore, the RBF clustering method combined with nearest neighbor clustering method and k-means clustering method is adopted in pixel classification. This method can readjust the location of the cluster centers and improve the effect of RBF fusion method.Transiently Chaotic Neural Network (TCNN) is used in wavelet image fusion. This paper adopts the weighted average strategy for the fusion of the wavelet transform coefficients. In traditional weighting method of image fusion strategy, the weighted coefficients were statically obtained by subjective factors and rigid formulas. The weighted coefficients are obtained through the dynamic optimization of chaotic neural network in this paper. Chaotic neural networks have been proved to be powerful tools for escaping from local minimum. The energy function is object fuction to be resolved. The TCNN outputs the weighting coefficient of every wavelet transform pixel when the energy function of the neural network has achieved the global minimum. At the same time, the region structural similarity value of every wavelet transform pixel gets the global maximum according to the relationship between the structural similarity and energy. The wavelet transform coefficients of the fused image are got by using the weighting coefficients. The advantage of the algorithm is that the weighting coefficient is obtained through the dynamic searching optimization of the structural similarity. Experiments show that the structural similarity values of the fusion images obtained by the proposed method are greater than the results obtained by the region energy method. The TCNN method improves the performance of the fusion image effectively. The TCNN method exploits a novel thinking for fusion. Furthermore, the TCNN method can also adopt other evaluative criteria. Through this way, the fused image based on any criterion is expected to obtain the optimum of this criterion. A group of optimums can be obtained by using different criteria. These optimums can be used as the reference in fusion strategies because the compared standard images are not always available for the fusion strategies.