Method of Parallel Computing Based on CUDA Image Fusion

Abstract

The invention discloses a method based on CUDA (Computer Unified Device Architecture) image fusion parallel computing. The method based on CUDA image fusion parallel computing includes the steps: S1, image segmentation: acquiring source images to be fused and respectively segmenting the source images, and then obtaining a plurality of sub images after segmentation; S2, parallel fusion: forming a plurality of groups of sub images to be fused through one-to-one correspondence of the sub images of each source image, and performing image fusion on each group of sub images to be fused parallelly through different CUDA cores to obtain a plurality of groups of fused sub images; and S3, image splicing: splicing the obtained each group of fused sub images through the step 2, and then obtaining the final fused image. The method based on CUDA image fusion parallel computing has the advantages of being simple in implementation method, being high in fusion efficiency and having good fusion effect.

Description

technical field [0001] The invention relates to the technical field of image fusion, in particular to a method for parallel computing based on CUDA image fusion. Background technique [0002] Multi-source image fusion is an information fusion technology that uses images as the research object. It fuses different images obtained by using different sensors for the same target or scene into one image, and the fused image can reflect the information of multiple original images. In this way, the multi-source image information can be used synergistically, so that different forms of information can complement each other and achieve the purpose of comprehensive description of the same target or scene, making it more suitable for visual perception or computer processing. Image fusion technology has been widely used in the fields of target detection, tracking and recognition, and situational perception. like figure 1 As shown, the image fusion process firstly transforms the image or...

AI Technical Summary

Problems solved by technology

[0003] Currently, mainstream image transformation algorithms for image fusion include transformation algorithms based on wavelet analysis, transformation algorithms based on multi-resolution non-subsampling theory, etc. Low, time-consuming fusion process and other issues, this problem has also become one of the main constraints in the actual use of image fusion

The multi-resolution analysis method represented by wavelet transform mainly uses the characteristic that the human eye is sensitive to the transformation of local contrast, and selects the most prominent features in multiple original images, such as edges and line segments, according to certain fusion rules. etc., and keep these features in the final composite image. Although this type of algorithm has good time-frequency localization characteristics, it needs to be completed by convolution, the computational complexity is high, and it does not have translation invariance, which is easy to lead to repetition. In order to make up for the above shortcomings, the multi-resolution non-subsampling transformation algorithm proposed, such as the non-subsampling contourlet transform algorithm (NSCT), divides the image transformation into two parts: multi-scale decomposition and direction decomposition. Each step is completed by a specific filter, but due to the introduction of a large number of two-dimensional convolution calculations in the operation process, the fusion process takes too long

In addition, the fusion effect of different fusion rules is also different. For fusion rules such as based on coupled pulse neural network (PCNN) and non-negative matrix factorization (NMF), it has a better processing effect, but it contains a large number of iterative operations, which further restricts the image quality. Fusion performance

Specifically, for the fusion of two images with a pixel size of 256*256, a typical fusion framework based on non-subsampled contourlet transform algorithm (NSCT) and non-negative matrix factorization (NMF) will take up to 188.23s, which seriously restricts The practical use of image fusion technology

[0004] In order to accelerate the process of image fusion, one method is to reduce the complexity of the algorithm, but at the same time it will lead to poor fusion effect; another method is to implement the algorithm in parallel, such as using DSP, FPGA and GPU, etc. Specialized hardware devices, in which GPU has excellent parallel computing capability and higher storage bandwidth, and its computing capability is much higher than that of CPU; CUDA (Computer Unified Device Architecture, Unified Computing Device Architecture) is a kind of device launched on the basis of traditional GPGPU The software and hardware system using GPU as a data parallel device can be developed with the help of common languages ​​such as C and C++, which greatly reduces the difficulty of GPU development, and at the same time uses the parallel computing capability of GPU to obtain very considerable acceleration performance

However, the current GPU-based image fusion framework has high requirements for the algorithm itself, and usually the fusion effect is poor. It can only be used for acceleration in a specific framework, that is, the acceleration process is limited to a single algorithm, and it is difficult to form a unified algorithm that is applicable to all algorithms. Acceleration method, which increases the workload and complexity of development, and is more limited in actual use

[0005] Chinese patent application CN105245841A discloses a panoramic video monitoring system using CUDA, which forms a panoramic video by splicing multiple video streams, but this method is limited to videos obtained by similar sensors and does not involve fusion processing of images from multiple sensors, and It is a mosaic algorithm with low algorithm complexity, and the complexity is much smaller than the computational complexity of image fusion

Images