An image processing method and device based on embedded gpu and convolution calculation

Abstract

The invention discloses an image processing method and device based on an embedded GPU and the convolution calculation. The method comprises the steps of optimizing the convolutional calculation in an SSD algorithm; performing matrix transformation on an input image by adopting the memory optimized convolution expansion; using a CUDA to parallelly process to form an intermediate matrix, meanwhile, adopting the convolution kernel matrix row and column expansion alignment, partitioning after the convolution kernel matrix expansion so as to reduce the memory overhead during operation, finally, adopting a highly optimized cuBLAS matrix multiplication function in a CUDA library to carry out convolution calculation parallel acceleration, and finally, combining and outputting the matrixes. According to the method provided by the invention, the memory overhead can be reduced, the performance of the algorithm is improved, the advantages of the GPU parallel control are brought into play, the matrix multiplication time is reduced, and the calculation efficiency is improved.

Description

technical field [0001] The present invention relates to the technical field of computer vision, and more specifically relates to an image processing method and device based on embedded GPU and convolution calculation. Background technique [0002] Since the introduction of convolutional neural networks in the ImageNet competition, computer vision technology has made great progress in the past few years, showing great performance in various fields such as image classification, pattern recognition, and multimedia compression. Among them, the SSD algorithm has been widely used. The SSD algorithm uniformly performs dense sampling at different positions of the picture. Different scales and aspect ratios can be used for sampling, and then the convolutional neural network is used to extract features and then directly perform classification and regression. The whole process Only one step is required, and the speed is faster than RCNN series algorithms. The SSD algorithm has been op...

AI Technical Summary

Problems solved by technology

[0003] Due to the complexity of the SSD detection algorithm process, when it is implemented on an embedded hardware platform, it needs to consume a large amount of storage and computing units, which puts forward higher requirements for the hardware platform

Because programming with hardware such as DSP and FPGA is very difficult compared to software development, on the other hand, many algorithms for software are difficult to implement with hardware, and the cycle of hardware development is longer and the cost is higher. GPU as processor

Embedded GPUs can execute programs concurrently and support deep learning CUDA libraries; however, embedded GPUs have limited memory, so how to optimize the memory utilization and runtime of convolution calculations on embedded platforms is very important for image processing

Images