A vision chip architecture for image recognition based on improved convolutional neural network

In this paper, we propose a novel architecture of vision chip based on CNN (convolutional neural network) and improve the traditional LeNet-S algorithm. From the results of the experiment and performance, the accuracy of the proposed chip with LeNet-S increased to 97.29%. According to the characteristic of deep pipeline among independent layers of CNN, a 3-stage pipeline CCE (Convolutional Computing Element) is designed and also introduce the logical data flow in CCE array in detail. Finally the data compression and decompression circuit are added to the vision system to reduce the bandwidth of off-chip memory which is the highest cost in data transmission. The storage reduction is nearly to 30%, depending on the image processed.