This paper presents a method for image retrieval using local grayvalue invariants computed at automatically detected interest points. The method is robust to partial visibility, similarity transformations, extraneous features, and small perspective deformations. The approach involves computing differential grayvalue invariants at interest points, which are invariant to rigid displacements and scale changes. A voting algorithm and semilocal constraints are used to ensure reliable retrieval. Indexing via a multidimensional hash-table allows efficient retrieval from a database of over 1,000 images. Interest points are local features with high informational content, identified using a detector that computes the auto-correlation function of image gradients. The local jet, derived from Gaussian derivatives, is used to compute differential invariants. These invariants are then used in a multiscale framework to ensure invariance to scale changes. A voting algorithm is used to handle mismatches and outliers, while semilocal constraints reduce the number of mismatches. Multiscale approach ensures robustness to scale changes, and the method is tested on a database of 1,020 images, showing high recognition rates even under varying conditions. The method is effective for retrieving images from large databases, even when only partial views or parts of the image are available. The approach is robust to image rotation, scale changes, viewpoint variations, and partial visibility. The method has been tested on a variety of images, including paintings, aerial images, and 3D objects, demonstrating its effectiveness in recognizing objects under different conditions. The results show that the method can achieve high recognition rates, even when only a small portion of the image is available. The method is also efficient, with retrieval times under five seconds for a database of 1,020 images. The approach is further enhanced by semilocal constraints, which improve the recognition rate by ensuring consistency in local features. The method is robust to statistical uncertainty and can handle a wide range of image variations.This paper presents a method for image retrieval using local grayvalue invariants computed at automatically detected interest points. The method is robust to partial visibility, similarity transformations, extraneous features, and small perspective deformations. The approach involves computing differential grayvalue invariants at interest points, which are invariant to rigid displacements and scale changes. A voting algorithm and semilocal constraints are used to ensure reliable retrieval. Indexing via a multidimensional hash-table allows efficient retrieval from a database of over 1,000 images. Interest points are local features with high informational content, identified using a detector that computes the auto-correlation function of image gradients. The local jet, derived from Gaussian derivatives, is used to compute differential invariants. These invariants are then used in a multiscale framework to ensure invariance to scale changes. A voting algorithm is used to handle mismatches and outliers, while semilocal constraints reduce the number of mismatches. Multiscale approach ensures robustness to scale changes, and the method is tested on a database of 1,020 images, showing high recognition rates even under varying conditions. The method is effective for retrieving images from large databases, even when only partial views or parts of the image are available. The approach is robust to image rotation, scale changes, viewpoint variations, and partial visibility. The method has been tested on a variety of images, including paintings, aerial images, and 3D objects, demonstrating its effectiveness in recognizing objects under different conditions. The results show that the method can achieve high recognition rates, even when only a small portion of the image is available. The method is also efficient, with retrieval times under five seconds for a database of 1,020 images. The approach is further enhanced by semilocal constraints, which improve the recognition rate by ensuring consistency in local features. The method is robust to statistical uncertainty and can handle a wide range of image variations.