Hardware Implementation of Optimized LeNet-5

Convolutional Neural Network (CNN), a feedforward neural network, is widely used in large-scale machine learning tasks such as image recognition, image classification, and object detection. Although CNN has strong performance, its application in daily embedded systems is limited due to the complex network hierarchy and a large amount of data. In this paper, the LeNet-5 network structure is optimized. The recognition accuracy of the optimized network on the MNIST dataset is 98.32%, but the number of weights is reduced by more than ten times compared with the LeNet-5 network, and data quantization is used to convert the weights into 8-bit fixed-point numbers, which is more suitable for deployment in the hardware with limited resources and power consumption. At the same time, different hardware structures and optimization strategies are designed for each layer of the network, and the convolution, pooling and fully connected layers are computed in parallel and optimized using methods such as loop unrolling and pipelining. Finally, the feasibility of the hardware design is verified by simulating each layer using the MNIST dataset and quantized weights.