This paper addresses the problem of retrieving images from large databases using local grayvalue invariants computed at automatically detected interest points. The method employs a voting algorithm and a local consistency criterion to enable efficient retrieval from a database of over 1,000 images. Experimental results demonstrate the effectiveness of the approach in handling partial visibility, similarity transformations, extraneous features, and small perspective deformations. The paper discusses existing recognition methods, including geometric and luminance-based approaches, and highlights the limitations of each. The proposed method uses local grayvalue features computed at interest points, which are robust to partial visibility and photometric information. The characterization is based on differential grayvalue invariants, ensuring invariance under rigid displacements and robustness to scale changes through a multiscale approach. The retrieval algorithm uses the Mahalanobis distance to quantify similarity and a voting mechanism to handle outliers and mismatches. Multidimensional indexing and semilocal constraints further enhance the accuracy and efficiency of the retrieval process. Experimental results show high recognition rates under various conditions, including image rotation, scale change, viewpoint variation, and partial visibility. The method is evaluated on a database containing 1,020 images, including paintings, aerial images, and 3D objects, achieving a recognition rate above 99%. The paper concludes by discussing potential extensions and improvements to the method.This paper addresses the problem of retrieving images from large databases using local grayvalue invariants computed at automatically detected interest points. The method employs a voting algorithm and a local consistency criterion to enable efficient retrieval from a database of over 1,000 images. Experimental results demonstrate the effectiveness of the approach in handling partial visibility, similarity transformations, extraneous features, and small perspective deformations. The paper discusses existing recognition methods, including geometric and luminance-based approaches, and highlights the limitations of each. The proposed method uses local grayvalue features computed at interest points, which are robust to partial visibility and photometric information. The characterization is based on differential grayvalue invariants, ensuring invariance under rigid displacements and robustness to scale changes through a multiscale approach. The retrieval algorithm uses the Mahalanobis distance to quantify similarity and a voting mechanism to handle outliers and mismatches. Multidimensional indexing and semilocal constraints further enhance the accuracy and efficiency of the retrieval process. Experimental results show high recognition rates under various conditions, including image rotation, scale change, viewpoint variation, and partial visibility. The method is evaluated on a database containing 1,020 images, including paintings, aerial images, and 3D objects, achieving a recognition rate above 99%. The paper concludes by discussing potential extensions and improvements to the method.