An embedded convolutional neural network acceleration method based on ARM

A convolutional neural network and convolutional neural technology, applied in the field of embedded convolutional neural network acceleration, can solve problems such as inefficiency, achieve the effect of wide use space and improve computing efficiency

CN109447239AActive Publication Date: 2019-03-08SOUTH CHINA UNIV OF TECH
2 Cites 4 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Publication Date
2019-03-08

Smart Images

  • Figure 1
    Figure 1
  • Figure 2
    Figure 2
  • Figure 3
    Figure 3
Patent Text Reader

Abstract

The invention discloses an embedded convolutional neural network acceleration method based on ARM, which overcomes the shortage of hardware resources of embedded equipment and the problem of high computational complexity of the convolutional neural network. The time-consuming convolution of 1*1 and 3*3 depth separable convolution are commonly used in lightweight convolution neural networks, and are optimized by using ARM NEON technique. In particular, the 1*1 convolution is first rearranged, then ARM NEON vector optimization is used to optimize the 3*3 depth separable convolution, which speedsup the computation of convolution neural network and makes full use of the hardware computation resources of embedded equipment, so that the convolution neural network deployed in the embedded terminal runs faster and more practical.
Need to check novelty before this filing date? Find Prior Art

Description

technical field

[0001] The present invention relates to the technical field of embedded convolutional neural network acceleration, in particular to an ARM-based embedded convolutional neural network acceleration method. Background technique

[0002] Deep learning algorithms based on convolutional neural networks have achieved great success in various fields of computer vision. However, with the continuous improvement of the performance of the deep convolutional neural network, the number of parameters of the network is increasing, and the amount of calculation is also becoming larger and larger. Due to the high requirements on hardware computing power of deep convolutional neural networks, it has become a challenge to deploy deep convolutional neural networks on devices with limited computing resources such as embedded devices.

[0003] At present, it has become a feasible method to design a lightweight convolutional neural network structure and deploy the structure to embe...

Examples

Embodiment Construction

[0035] The optimization method of the present invention will be further described in detail in conjunction with the drawings and MobileNetV1 below, but the present invention is also applicable to other neural networks using 1×1 convolution and 3×3 depth separable convolution.

[0036] like image 3 As shown, the ARM-based embedded convolutional neural network acceleration method provided by the present invention comprises the following steps:

[0037] Step 1, use Caffe or other deep learning frameworks to train the lightweight convolutional neural network MobileNetV1.

[0038] Step 2, export the trained MobileNetV1 network structure and weights to a file.

[0039] Step 3, the design program imports the weight file, and realizes the forward calculation of the neural network according to the trained network structure. Different layers in the neural network can be represented by different functions. Function parameters include layer specification parameters, input feature maps,...