Research On Several Issues About Image Processing Based On Pulse Coupled Neural Networks

From the 90s 20th century, Eckhorn and others made a research on the burst of visual cortical neurons of oscillation of the cats and monkeys, and then established the model of PCNN (Pulse coupled neural network, PCNN) with the following further research.PCNN is a simplified model of the mammalian cerebral cortex neurons of the visual center, and the PCNN has many excellent characteristics in the image processing, such as scale invariance, rotation invariance, intensity invariant, and distorted invariant features, at the same time, PCNN also can be combined with other classic image processing methods, such as median filter, Wiener filter and some classic a pattern classifier such as BP, Bayesian, wavelet, to a certain extent make up for their own shortcomings.At present, PCNN is still a lot of work to do, such as adaptive adjustment parameters and the effect of image processing is a hot and difficult research at home and abroad, the effect of PCNN image processing is not a certain standard to measure, can only be determined by subjective judgment, so get a good evaluation standard is very necessary.The main research of the thesis is the models, parameters adjustment and the application in image processing, which mainly focus on the feature extraction and image recognition. As is known, image recognition is divided into four stages, witch are image acquisition, preprocessing, feature extraction and image recognition. The last three stages is the emphasis in this article, and then implement the entire process through simulation experiments. In the preprocessing stage, it’s necessary to remove the noise in the image and make some necessary geometric transformation, and then make a brief introduction for image denoising in this article, for example, the mean filtering and the median filtering algorithms, and also given the results by experiments. During the feature extraction stage, it’s mainly introduce the feature extraction about the simple and complex graphics respectively, the main features involved in the article are time series, entropy sequence and the Euler sequence. In the pattern recognition stage, the processing is mainly attributing to the pattern classifier, such as the minimum distance classifier, the BP neural network classifier, and so on.The recognition ratio of the simulation results has proved the ability of the feature extraction.The main section of the paper I am such arrangement:First of all, the first chapter,a simple description and summary is given about the study status of PCNN, the aim of the research, the research significance, the main work and the organized are described respectively in detailed.Secondly, in the second chapter, we introduce the basic knowledge and principle. of neural network.In the third chapter, I give a detailed derivation and description of the theory and model of PCNN, introduces the source of the PCNN neuron model, working principle and characteristics of the neurons firing cycle, and also discussed the pulse coupled neural network model in the coupled and uncoupled two kinds of circumstances neurons work principle, and combined with the circuit model supplementary explanation. Specifically, the pulse coupled neural network and pixel image is one one correspondence, that is to say, an image has many pixels, then the network will have the number of neurons and the corresponding. In neurons without coupling connection, each neuron mutually independent, external input is greater, the smaller neurons firing cycle, the ignition frequency is higher, this shows that, neuron is not affected by outside circumstances, the release of pulse frequency depends on the size of the pixel gray value. While in the practical application, most of the coupling model, reason is connected to the input to the feedback input is the key to transfer information between neurons. In the coupled case, the nature of its neurons due to coupling with similar or gray space capture similar neurons and the ignition is similar to PCNN in practical applications has similar or similar gray space and release mechanism of synchronous pulse. At the same time, by adjusting the connection coefficient can enlarge or shrink the neuron connection strength, thereby affecting neuron firing size area coverage.In the fourth chapter, study the application of PCNN in image processing. Mainly focus on the image preprocessing, and the principle, pattern classifier algorithm are discussed, and the simulation of the experimental process.We know that, in practical application, for a involved in the recognition of image, the pretreatment in prophase work is very important, the quality of the pre processing good or not will affect the recognition accuracy of the results. Usually the image can be various kinds of noise pollution, and the noise may be generated in the process of image acquisition or transmission, may also be in the image produced by the quantization process, in order to improve the recognition rate, the denoising, ensure to the maximum extent of removing noise and does not destroy the the original image information, the contour and the lines are not damaged, smoothing filtering methods are commonly used spatial filter and frequency domain filter.Cause of the noise is decided by the distribution characteristics of noise in an image pixel and image signal. We use the linear filter, median filter denoising, using a minimum distance classifier and BP neural network classifier for pattern recognition. Good effect.The fifth chapter is the focus of this article. I aimed at the feature extraction and do some work. If the PCNN is used for simple geometry, a character image feature extraction and recognition, and on the basis of previous research to improve image recognition system based on PCNN, including the image acquisition, preprocessing, feature extraction and recognition of the whole process, and give the detailed theory is introduced, the experimental process and experimental results. Application of denoising and color morphology.Johnson has been demonstrated by the experimental results of each input image has its own characteristic time sequence only, but also by numerous experiments show that most of the input image has its own entropy sequence only, so this paper according to the two characteristics, combined with the characteristics of PCNN itself in the field of image processing (with rotation, scale, translation and twisted invariance in the feature extraction) to identify the input images, experimental results show that PCNN can successfully identify the image after adding noise in image recognition. We can also use PCNN in image processing advantages, combined with the characteristics of the image itself, namely each image has its own characteristic time sequence the only and most images have the information entropy of their own, so the combination of these two unique characteristics of image recognition can be used as basis, through the experimental results show that, on the same side image, after adding different noise processing, the characteristics of time series and entropy sequence is then obtained by the PCNN, do the Euclidean distance judgment on it, and finally the recognition results obtained, if the characteristics of time series and entropy sequence of two images are very close, can be considered the same image.The key of image recognition is how to extract the main features of the image better, and then through a certain criterion to identify. Through the experiment, this paper introduce the Euler number feature sequence as the topological features of image recognition, and to identify the characteristics of time series and the entropy sequence characteristics of the results, obtained the ideal effect in the experiment, show that the method and the model is effective in image recognition, and fast, open up a new the way for image recognition.Based on the application of morphological filters omnibearing multi angle structure compound in a gray image, this paper proposes a color morphological filter for all directions based on structure element composite. This algorithm preserves the gray image of the advantages, and based on the results of optimization, the algorithm finally get the ideal denoising effect. Specific performance, based on gray level and two value morphology, the morphology is applied to color images, introduces the model of color image and model, put forward the color morphological transformations in space, and on the basis of previous research results, in space, put forward a full directions composite color scale morphological filter,. But the effect of this algorithm to obtain the graph can be seen only on a noise removal, the treatment effect is not obvious, so we in the original basis of the two kinds of experimental effect diagram according to the reality of certain provisions of the implementation of weighted stack, can also be two remove noise and preserve the image edges, but from the graph (f) that the light noise and dark noise also do a deal, but the image or noise small, we must be on the basis of further improvement, in order to obtain the better effect.Then, in the sixth chapter,the setting of the parameters for PCNN model were studied and discussed, the entropy information was used to determine the best number of iterations and the best threshold for segmentation, and then which can also determine the dynamic threshold of the time decay constant, and obtained good results through the experiments.Entropy is the statistical characteristics of image information to reflect the information contained in document size, the image segmentation is a method to estimate PCNN model proposed by time decay parameter estimation method, is proposed for information histogram analysis, because of the time decay constant control iterations neurons in a cycle the. If an image is a gray level, the total amount of information statistics all the gray level, according to the uncertainty principle, the equal probability of each layer gray image appears, the maximum information entropy. Through the statistics histogram to determine the decay time constant and achieve better segmentation effect, and through the experience value determines the PCNN decay time constant part segmentation effect is not ideal,The information entropy is the reaction of an image contained in the amount of information statistics, because the PCNN model itself is not iterative one output a two value image, the two value image is only 0 and 1 in two cases, and 0 and 1 in the two values in the image reflects the size of the information of the original image, through the output of two value image to extract the information entropy. Because each iteration in PCNN output of the two value image change, as the number of iterations and therefore, the information entropy changes, when the information entropy to a maximum value, two most information of the original image contains the moment corresponding, according to the principle to determine the optimal number of iterations.Through the discussion of the PCNN model of dynamic threshold decay time constant and optimal iteration times, as can be seen, the information entropy plays an important role in image processing. The maximum information entropy to determine the optimal number of iterations of the PCNN model are used in many fields of image processing, information entropy reflects an image pixel distribution. This chapter is discussed the parameters of PCNN model based on literature documents and selection, and the corresponding experimental results show its effectiveness, to determine the parameters of PCNN model is always the emphasis and difficulty of the research, and therefore subject to the concern of many scholars, such as PCNN, literature and traditional back-propagation BP neural network combined, this paper proposes an adaptive pulse coupled neural network, and the adaptive parameter for the next step of work to do.Finally, the main work and innovation in this thesis are given a brief description, for example, the Euler sequence was selected as the feature, Presents a color morphological filter for all directions based on structure element composite. This algorithm preserves the gray image of the advantages, and based on the results of optimization, the algorithm finally get the ideal denoising effect. At the end of the parameters of PCNN were studied and discussed, the information entropy is used to determine the number of iterations and the optimal segmentation threshold, so as to further determine the attenuation constant dynamic threshold PCNN model of the time, and through the experiment indicates that the method to determine the time decay parameters obtained good results,and so on, and the further research directions are showed in the last part.