金字塔模板匹配算法融合NMSFast以及优化研究

Integration of Pyramid Template Matching Algorithm with NMSFast and Optimization Techniques

摘要: 图像模板匹配是计算机视觉领域的一项重要任务，它在许多应用中都有广泛的应用。然而，传统的模板匹配算法在大规模图像和复杂场景下存在计算量大、效率低的问题。为了解决这些问题，本文提出融合快速非最大抑制(NMSFast)的金字塔模板匹配算法，提高准确度，并通过特征提取、查表优化、OpenMP并行、量化等技术对其优化，从而提高效率。基于Sobel获取图像的梯度信息，并结合阈值和强度条件来筛选候选特征点以达到特征提取。通过查表创建模板特征和对应搜索图像特征之间的关联关系和缩放因子和旋转角度对应的变换矩阵的索引表。将特征数据进行量化，其转换为更简单的浮点数，对角度图像进行8方向量化，结合阈值过滤无效角度值。以上优化能够减少计算量和存储空间的消耗。OpenMP并行技术对金字塔进行并行分层搜索，将单线程变成多线程，可以提高算法的运行速度。实验结果表明，所提出的金字塔模板匹配算法融合NMSFast算法在大规模图像匹配任务中，运算时间提高51%，精度提高1.7%。

Abstract: Template matching in image processing is an important task in the field of computer vision and has wide applications in various domains. However, traditional template matching algorithms suffer from high computational complexity and low efficiency, especially when dealing with large-scale images and complex scenes. To address these issues, this paper proposes a pyramid template matching algorithm that integrates the NMSFast (Non-Maximum Suppression Fast) al-gorithm for improved accuracy. Several optimization techniques, including feature extraction, table lookup optimization, OpenMP parallel programming, and quantization, are employed to enhance the efficiency of the algorithm. The proposed algorithm utilizes the Sobel operator to extract gradient information from images and applies thresholding and intensity conditions to filter candidate feature points during feature extraction. Table lookup is used to establish the correspondence between template features and the corresponding features in the search image, along with indexing tables for scaling factors and rotation angles. Feature data are quantized to simplify representation, converting them into floating-point numbers. The angle image is quantized into 8 directions, and invalid angle values are filtered out using thresholding. These optimizations significantly reduce computational complexity and storage requirements. The OpenMP parallel programming technique is employed to perform parallel hierarchical searches on the pyramid, transforming the algorithm from single-threaded to multi-threaded, thereby improving its runtime performance. Experimental results demonstrate that the proposed pyramid template matching algorithm integrated with NMSFast achieves a 51% reduction in computation time and a 1.7% improvement in accuracy for large-scale image matching tasks.