What are the methods of image matching and recognition based on local features

Features extracted from an image can form a vector, and the degree of similarity between two images can be calculated by defining a distance or similarity measure.

Feature matching is a key part of image retrieval, characterized by feature dependency, and different features should be used with different metrics. In the process of retrieval, according to the algorithm of the system similarity metrics to calculate the degree of similarity between the query features and each group of features corresponding to the feature library, and then sort the results from largest to smallest to get a sequence of matching images returned to the user. In the meantime, through human-computer interaction, the retrieved results can be refined step by step, and the scope of the matching set can be narrowed down continuously, so as to locate the target. The matching process often utilizes the distance function between feature vectors for similarity metrics, which mimics the human cognitive process and approximates the cognitive ordering of the database. Commonly used distance metric formulas are: Minkkowsky distance, Manhattan distance, Euclidean distance, weighted Euclidean distance, Chebyshev distance, Mahalanobis distance, etc.

The Manhattan distance is simple to calculate, and the effect is better, and it is widely used; the weighted Euclidean distance takes into account the importance of different components, and it is also more commonly used; the Mahalanobis distance takes into account the statistical characteristics of the samples and the correlation between the samples, and it is often used in the cluster analysis. When the integrated features are used for retrieval, the feature vectors need to be normalized to make the integrated features of the eigenvectors have the same status in the similarity distance calculation.