Image Processing Based On Pulse Coupled Neural Network

Pulse Coupled Neural Network(PCNN),called the third generation artificial neural network,is derived from the research on the information processing mechanism of mammal visual cortex.It has been widely applied to the field of image processing with its features of neuron synchronous burst and capture.However,the traditional PCNN has a complex structure and too many parameters.What's worse,these parameters often need to be manually set or rely on empirical formulas.This often makes the model performance unstable.Thus,it is always hard to achieve the image processing effect that people need.Based on the information processing mechanism of biological neurons,a modified model is proposed,which inherits the traditional PCNN such as pulse synchronization of neurons and the capture property.It also incorporates the Hebb rule to bring it closer to the characteristics of real biological neurons.It is applied to image segmentation,image retrieval,face recognition and other image processing problems,and have obtained a relative good effect.In detail,the following three aspects are studied:1.Aiming at the complex structure and too many parameters of traditional PCNN,this paper proposes a modified model that removes the self-feedback input of feeding input and linking input,and only retains the linking modulation of linking input.It also changes the threshold function into an exponentially decaying form monotonically.So the modified model is much simpler than the traditional PCNN.What's more,the Hebb rule is also introduced to our proposed model.As a result,the proposed model can improve the updating method of weight coefficients and feeding input,since the Hebb rule is more similar to real biological neuron.According to its experimental results in image segmentation and face recognition,it is proved that it has a good effect in these areas.2.For specific image segmentation tasks,the threshold is initialized by the Otsu algorithm,and as an iterative termination condition,so that the modified model can automatically determine the best segmentation result.For the isolated noise points that may appear before and after the image input,a single-step iterative optimization model is proposed to make the edges clearer and improve the image segmentation quality to some extent.Compared with the traditional PCNN and Otsu algorithm,from the experiment results,it is shown that the modified model can provide attractive performance.3.From the experiments,it is proved that the time series in the modified model has the characteristics of rotation and distortion invariance in a certain degree.This property makes the modified model be used to extract image features.Hence,the modified model is further applied to the problem of image retrieval.Aiming at image retrieval problem of the same object,the k-Nearest Neighbors-based retrieval method is proposed.Simulation results show the efficiency of our modified model.Then we utilize the face recognition problem as an example to deal with the problem of the same category of image retrieval.We develop a recognition method based on k-Nearest Neighbor and ensemble learning.It is proved through experiments that it has a higher recognition rate in face recognition.