Optimization Technology of CNN Based on DSP

Convolution neural network has important applications in the field of image recognition and retrieval, face recognition and object detection in deep learning. In the training of convolution neural network, 2D convolution, spatial pooling, linear mapping and other operations of forward propagation will have a huge computational complexity. In this paper, an effective optimization technique is proposed to map the convolutional neural network to the digital processor DSP. These technologies include: fixed-point conversion, data reorganization, weight deployment and LUT (look-up table). These technologies enable us to optimize the use of resources on the C66x DSP. The experiment is carried out on Texas Instruments C6678 development board, and the optimization technique proposed in this paper can be applied to multiple open-source network topologies.