Method and device for comparing similarities of high dimensional features of images

1. A computer-implemented method for comparing similarities of high dimensional features of images, characterized in comprising:

reducing dimensions of extracted eigenvectors of the images by Locality Sensitive HASH (LSH) algorithm to obtain reduced dimensional eigenvectors;

averagely segmenting the reduced dimensional eigenvectors to obtain segmented reduced dimensional eigenvectors and establishing a segment index table for the segmented reduced dimensional eigenvectors;

retrieving the segmented reduced dimensional eigenvector of a queried image from the segment index table to obtain a candidate sample set;

performing a similarity metric between each sample in the candidate sample set and the reduced dimensional eigenvector of the queried image to select at least a subset of samples in the candidate sample set as similar images for the queried image; and

providing the similar images as a response for the queried image.

2. The method according to claim 1 , characterized in that the image eigenvectors are extracted with a neural network constructed using a deep learning technique.

3. The method according to claim 2 , characterized in that the neural network is a convolutional neural network.

4. The method according to claim 1 , characterized in that prior to the step of averagely segmenting the reduced dimensional eigenvectors, the method further comprises:

experimentally determining an optimal segment length on a smaller verification set.

5. The method according to claim 1 , characterized in that the step of averagely segmenting the reduced dimensional eigenvectors and establishing a segment index table comprises:

averagely segmenting the reduced dimensional eigenvectors, using the segmented eigenvectors as index items, and calculating fingerprint of each of the index items;

performing a remainder operation on the fingerprint with a prime number which is closest to a predetermined number of entries contained in the segment index table to obtain entry addresses for the index items; and

inserting the reduced dimensional eigenvectors into the segment index table according to the obtained entry addresses to establish the segment index table.

6. The method according to claim 1 , characterized in that the step of retrieving the segmented reduced dimensional eigenvector of a queried image from the segment index table to obtain a candidate sample set comprises:

accessing an entry address of the segmented reduced dimensional eigenvector of the queried image to obtain a conflict set;

extracting the reduced dimensional eigenvectors corresponding to a node of the conflict set, which has the same fingerprint as that of the segmented reduced dimensional eigenvector of the queried image, as a candidate set; and

combining the candidate set obtained by respective segment retrievals and removing therefrom duplicated reduced dimensional eigenvectors to obtain a candidate sample set.

7. The method according to claim 1 , characterized in that the step of performing a similarity metric between each sample in the candidate sample set and the reduced dimensional eigenvector of the queried image comprises:

calculating scores of Manhattan distances between each sample in the candidate sample set and the reduced dimensional eigenvector of the queried image;

sorting the scores according to an ascending order, and taking images corresponding to the samples with a predetermined number of top scores as the similar images of the queried image.