Due to the lighting, translation, scaling and rotation, image matching is a challenge task in computer vision area. In the past decades, local descriptors (e.g. SIFT, SURF and HOG, etc.) and global features (e.g. HSV, CNN, etc.) play a vital role for this task. However, most image matching methods are based on the whole image, i.e., matching the entire image directly base on some image representation methods (e.g. BoW, VLAD and deep learning, etc.). In most situations, this idea is simple and effective, but we recognize that a robust image matching can be realized based on sub-images. Thus, a block-based image matching algorithm is proposed in this paper. First, a new local composite descriptor is proposed, which combines the advantages of local gradient and color features with spatial information. Then, VLAD method is used to encode the proposed composite descriptors in one block, and block-CNN feature is extracted at the same time. Second, a block-based similarity metric is proposed for similarity calculation of two images. Finally, the proposed methods are verified on several benchmark datasets. Compared with other methods, experimental results show that our method achieves better performance.

由于照明，平移，缩放和旋转，图像匹配在计算机视觉领域是一项艰巨的任务。在过去的几十年中，本地描述符（例如SIFT，SURF和HOG等）和全局特征（例如HSV，CNN等）在此任务中起着至关重要的作用。但是，大多数图像匹配方法都是基于整个图像，即直接基于某些图像表示方法（例如BoW，VLAD和深度学习等）匹配整个图像。在大多数情况下，这种想法简单有效，但是我们认识到可以基于子图像实现鲁棒的图像匹配。因此，本文提出了一种基于块的图像匹配算法。首先，提出了一种新的局部合成描述符，该描述符结合了局部梯度和颜色特征与空间信息的优势。然后，VLAD方法用于在一个块中编码所提出的复合描述符，并同时提取块CNN特征。其次，提出了基于块的相似度度量用于两个图像的相似度计算。最后，在几种基准数据集上验证了所提出的方法。与其他方法相比，实验结果表明我们的方法具有更好的性能。