PCNN-based Deep Models For Image Classification And Segmentation

With the development of the internet era,a large number of images and videos on video websites and social networking sites have become an important part of people's daily information exchange.The way to effectively deal with the increasing image data has become an urgent problem in the field of computer vision.Computer vision refers to the use of electronic equipment instead of biological vision,and computer programs to simulate biological brains to process and analyze images and videos.In recent years,with the rise of deep learning,a variety of deep models have been applied to computer vision,achieving breakthroughs in accuracy and processing speed.Pulse Coupled Neural Network(PCNN)is a neural network model built based on simulating the visual physiological system of advanced mammals.Before the rise of deep learning,PCNN played a significant role in various aspects of image processing.Nowadays,however,it is difficult for a single PCNN model to meet the complex computer vision needs.Therefore,this paper,we explore the potential methods and technologies of PCNN-based deep neural models in computer vision applications,particularly on natural image and biomedical image processing.The main innovations in this article are as follows:(1)There is an essential principles in the PCNN model,i.e.the higher the neuron strength,the stronger its internal activity,and the earlier its fire time and the synchronization pulse issuance.Combining with the simplified PCNN(SPCNN)model and the spiking deep neural network(SDNN),we propose an adaptive,unsupervised,spiking communication,SPCNN-SDNN model based on SPCNN temporal coding for image classification.The SPCNN-SDNN model uses the STDP unsupervised training algorithm and introduces the SPCNN temporal coding layer.SPCNN temporal coding can obtain adaptive spiking-timing maps based on different input images.Experiments show that under the same experimental conditions,our proposed SPCNN-SDNN model can achieve higher classification accuracy on the Caltech face/motorbike and MNIST datasets.(2)Since the PCNN model is efficient in simple image processing,we propose a new nuclear segmentation method based on PCNN+U-Net.Specifically,we introduce an image enhancement by PCNN into the original U-Net network as an image preprocess layer.Then,we adjust the structure of the original U-Net model:replacing the valid convolution with the same convolution;substituting the two-channel feature map of the original U-Net network output with a single-channel feature map.Experimental results show that the segmentation effect of this method on the MICCAI2017 data set is better than that of the U-Net model.(3)To further explore the effect of combining the PCNN model and deep model in image process,we propose a new method for the nuclear segmentation of adherent cells based on PCNN+Caps-Unet.In this method,PCNN and Caps modules are introduced into the U-Net model to perform the nuclear segmentation of adherent cells.Experimental results show that compared with the original U-Net and Caps-Unet models,the method presented in this paper is better than that of the original U-Net and Caps-Unet models.