Research On Application Of Cellular Neural Networks In Image Processing

Cellular Neural Networks (CNN) is a large-scale nonlinear array processor doing real-time signal procession, of which the local interconnection feature makes it suitable for VLSI (very large scale integrated) implementation. It is a multidisciplinary subject and hot topic in the neural network area with various applications in image and video signal processing.Practical researched and applications on CNN for image processing application are covered in this paper. CNN's ability to perform nonlinear, high-speed parallel computation and its advantage for hardware implementation finds its use in real-world applications. Cooperating with the morphology, regularization theory, and Markov Random Field technique, CNN is capable to do complex computation such as the image processing and ill-posed problems.There are four major contributions in this paper:1.Due to the importance of the stability to a system, the dynamic range and stability of the CNN are analyzed from the aspect of its mathematical model and physical structure. Furthermore, the principle for its usage in image processing and the module designing strategy is covered.2.Two CNN based algorithms are introduced to do interference-stripes image processing, which also combines the conventional filtering, binary subdividing and stripe trimming techniques. The first algorithm proposed employs CNN adaptive filtering algorithm to combine the conventional filtering and binary dividing process together, which greatly simplifies the procedure of interference-stripes image processing and can select filter parameters adaptively according to the noise character of the image.The other algorithm which combines CNN and mathematics morphology is proposed for application of thinning and trimming the interference-stripes.The two algorithm can be unified by using same CNN system with different parameter sets or modules, providing a new method for interference-stripes and other optical image processing3. Shape from Shading (SFS) is a difficult problem in 3-D shape reconstruction.Treating SFS as an optimization problem and optimizing energy function of CNN, a 3-D shape can be reconstructed from 2-D images.The characteristics of parallel computation and easily hardware implantation make it an ideal method for real-time target recognition with SFS technology. The construction of energy function, the mapping relationship from SFS to CNN is introduced in details in this sector.4.As a key part of video compression, motion estimation is usually used to reduce inter-frame redundancy. After analyzing the similarity between CNN model and Markov Random Field(MRF). A new CNN based real time video motion estimation algorithm which combines the CNN and MRF parameter estimation with Bayesian theory and take the discontinuity point problem into consideration. This algorithm proposed not only has MRF's advantage of constructing local restriction condition but also has CNN's ability for hardware implantation and high-speed computation, thus a new and novel method for real-time video encoding.