Compression acceleration method of deep convolutional neural network for target detection

Abstract

The invention discloses a compression acceleration method of a deep convolutional neural network for target detection. The method comprises the following steps: constructing and training a deep convolutional neural network for target detection; carrying out quantitative test on all weight values in the deep convolutional neural network and activation values of all layers except the last layer after passing through an activation function, testing the detection performance loss condition of the network from small to large in quantitative step size, and selecting the maximum quantitative step size in a set loss range; determining a truncation range of a weight value and an activation value in the neural network by utilizing the quantization step length, limiting the neural network and training the network; and truncating and quantifying the deep convolutional neural network, and compiling a forward code. According to the method, the quantization technology is adopted to reduce the networkstorage capacity, 32-bit floating-point number operation in the network is converted into 8-bit integer operation, and meanwhile, the sparsity of the network is utilized to convert the layer meetingthe sparsity condition in the network into sparse matrix operation, so that the purpose of compressing and accelerating the deep convolutional neural network is achieved.

Description

technical field [0001] The invention relates to the fields of deep learning and artificial intelligence, in particular to a compression acceleration method of a deep convolutional neural network for target detection. Background technique [0002] Object detection has always been an important research direction in the field of computer vision due to its broad application prospects in information retrieval, autonomous driving, robot navigation, and augmented reality, and has received extensive research and attention from academia and industry. The traditional object detection system mainly uses some artificially designed features, such as Haar features and HoG features, etc., and uses classifiers such as support vector machines to classify images with sliding windows to achieve the effect of object detection. In recent years, with the rise of deep learning, convolutional neural networks have brought extremely effective solutions to target detection. The results obtained by met...

AI Technical Summary

Problems solved by technology

[0003] However, because the convolutional neural network is accompanied by a large amount of storage and calculation, for example, the classic VGG16 network requires about 520MB of storage and 15.3 billion multiplication and addition operations. Even some existing lightweight networks still require dozens of Megabytes of storage and millions of multiplication and addition operations, such a huge amount of storage and calculations limit the application of convolutional neural networks, especially in mobile devices and embedded devices

Images