Design and realization of deep learning coprocessor oriented to image recognition

In order to explore efficient design of the dedicated deep learning processor oriented to image recognition, this paper presents a deep convolutional neural network coprocessor which is applied in embedded image recognition system or SoC. In particular, the calculation of nonlinear active function and local response normalization are designed in PEs of the coprocessor based on approximate calculation so that the parallel computing in multiple PEs could be used instead of traditional serial computing based on additional ALU. In order to verify the function and performance of the coprocessor, two representative convolutional networks LeNet-5 and AlexNet are deployed on coprocessor respectively, and two datasets MNIST and CIFAR10 are used for training and testing. The prototype system of the coprocessor is implemented on XC6VLX240T of Xilinx, and the peak performance is up to 19.2GOP/s when it operates at 100MHz. The experiment results show that the test accuracy of the coprocessor is consistent with the deep learning framework Caffe, and the design of coprocessor meets the requirements of image recognition.