Multi-granularity parallel optimization method based on sequence image Harris-DOG feature extraction

Abstract

The present invention relates to a multi-granularity parallel optimization method based on sequence image Harris-DOG feature extraction. The method is characterized in that firstly the corresponding tasks are divided, a host terminal (a CPU) adopts a multi-thread mode to process, and an image is read and uploaded to a GPU global memorizer; then the CPU initializes the relevant parameters and uploads the parameter values to a GPU when carrying out the Harris feature extraction; the CPU calculates a Gauss template and binds to a GPU constant memorizer when carrying out the DOG feature extraction; an equipment terminal (the GPU) adopts a CUDA parallel mode to process, aiming at the same original image input, a Harris-DOG operator feature extraction algorithm utilizes the Harris operators and the DOG operators to extract the feature points respectively, and a calculation result is returned to the CPU; and finally, the host terminal constructs and merges the feature point sets. The multi-granularity parallel optimization method of the present invention enables the sequence image Harris-DOG feature extraction to be accelerated and optimized effectively.

Description

technical field [0001] The invention relates to CPU_GPU multi-granularity parallelism, CUDA, Harris-DOG feature extraction and image processing in computer science, especially a multi-granularity parallel optimization method based on sequence image Harris-DOG feature extraction. Background technique [0002] CUDA (Unified Computing Device Architecture) is a GPU-based general-purpose parallel computing architecture launched by NVIDIA. The CUDA architecture adopts the CPU+GPU heterogeneous model. The CPU is responsible for complex logic processing and transaction processing, and the GPU is responsible for computing-intensive large-scale data parallel computing. [0003] Such as figure 1 As shown, the CUDA architecture provides 6 types of device-side memory and 2 types of host-side memory. Shared memory is a readable and writable high-speed memory in the GPU chip, which can be accessed by all threads in the same block. The global memory is located in the video memory, and bo...

AI Technical Summary

Problems solved by technology

For the Harris-DOG feature extraction of sequence images, the number of images to be processed is huge. If only GPU acceleration is used, the execution time after acceleration is still long, which will cause the CPU to wait for a long time and waste the computing power of the multi-core CPU.

Images