Research On Spiking Neural Networks And Its Application On Image Retrieval

As the artificial neural network with the highest bionic nature, the spiking neural networks(SNNs) are proposed to simulate the biological brain function. It is also known as the third generation of the artificial neural network. Its model uses the spike time encoding method to represent and transmit information, which is more similar to the biological neurons than the first two generation of traditional neural networks. The recent researches also indicate that with this information processing method, the spiking neural networks possess higher capability both in information representation and computation. Consequently, the SNNs have aroused widespread concern and attention by researchers at home and abroad. By now, a lot of achievements have been made, however, the relative achievements on real world applications are still in the early stage. In the current internet age, more and more information in our society is transmitted by images or image related methods. Since the last century, the field of automatic identification and retrieval of image recognition are becoming increasingly urgent requirements. Then, applying the spiking neural networks to the image processing applications is a very important study. It can take spiking neural networks to more applications. Besides, it can motivate the rapid development in the image processing field.To apply the spiking neural networks successfully to the image retrieval system, the main work of this thesis is contents of three parts:(1) we propose a method based on the spiking neural networks to detect the image edge. Our new method combines the advantage of the temporal information representation of the spiking neural networks and the advantage of the spatial network connection method by the convolution neural networks, and achieves a good performance on image edge detection.(2) we propose a new method based on the spiking neural networks to locate the angular feature points of the image. It employs the template to achieve angular points feature location. To two types of images with edge detected or grayscale image, our method has two variations to complete this angular point detection.(3) we design a spiking neural network for the image retrieval in our thesis. This method combines the low-level visual features of the image and the semantic feature of the image, which makes our method more comprehensive than traditional methods employing only one type of feature. The low-level visual features are represented by the angular feature points detected above and the introduced feature obtained by the spiking based convolution. The semantic feature of the image is represented by the label information of the image. By training a supervised spiking neural network, we can predict the label of a retrieval image accurately. To test the availability of these proposed methods, some simulations are conducted on the Matlab platform. The simulation results prove that our method based on the Spiking neural networks can be applied to the image processing problems successfully.