Algorithm Modeling Of Visual Neural Networks And Its System-on-Chip Implementation

As the bio-inspired features modeling of biological neural networks,visual neural networks have achieved many successes on various applications of visual recognition and object detection.Therefore,it is important and realistic to do algorithm research and architecture design on related aspects of visual neural networks,including network structure,information transmission,network training and inference.In this paper,related researches have been done on deep neural networks,cellular neural networks,and spiking neural networks,respectively.The related algorithms have been proved theoretically and verified experimentally in the visual applications of cross-domain object detection,image filtering,visual illusion analysis,object localization and etc.In addition,a customized instruction-set is designed for the computational implementation of the network inference.As a result,the central processing unit(CPU)of realizing software optimization and the neural network processing unit(NPU)of realizing hardware optimization are integrated efficiently in the same single-chip heterogeneous system.The main research contents are listed as below:(1)Propose an algorithm to decrease the negative influence of domain divergence for the training of deep neural networks.The algorithm is based on the combination of domain transfer learning and deep-feature extracting.Faster R-CNN and YOLOv2 are selected to represent the applied deep neural networks in the field of object detection.Accordingly for cross-domain object detection,we proposed two joint network architectures which are respectively based on the Faster R-CNN and YOLOv2.(2)Propose several models of cellular neural network applied in image filtering and visual illusion analysis.And propose a model of spiking neural network for object localization.The models of cellular neural network are inspired from the structure of biological neural networks,and the model of spiking neural network is inspired from the mechanism of information transmission in biological neural networks.(3)Propose a hardware architecture of layer computing,which is based on hardware optimization and used for the visual neural networks with the characteristic of convolution computation.And propose an on-chip memory system to reuse the data of layer computing and collaborate with the hardware architecture.In addition,a customized instructionset is proposed as a cooperative bridge between software optimization and hardware optimization.As a result,a system-on-chip(SoC)implementation of visual neural networks has been completed in the experiment.Finally,The efficiencies of the proposed algorithms and hardware architectures have been verified experimentally on graphics processing unit(GPU)server and SoC platform.