Conditional similarity-based image identification and matching with reduced labels

1 . A system for matching one or more images using conditional similarity pseudo-labels, comprising:

a non-transitory computer-readable medium operatively coupled to a processor device, the processor being configured for:

analyzing an unlabeled dataset of images;

accessing a foundational vision-language model trained on a plurality of image-text pairs;

defining a set of attributes each including multiple possible values for generating pseudo-labels based on notions of similarity (NoS);

generating text prompts for each attribute value using a prompt template and encoding the text prompts using a text encoder of the foundational model;

processing each image in the unlabeled dataset of images through a vision encoder of the foundational model to obtain visual features;

comparing the visual features against encoded text prompts to assign a pseudo-label for each attribute for each image; and

training a Conditional Similarity Network (CSN) with the pseudo-labeled images to generate a conditional similarity model.

2 . The system of claim 1 , wherein the foundational vision-language model is Contrastive Language-Image Pre-Training (CLIP).

3 . The system of claim 1 , wherein the set of attributes includes specific user-defined conditions and criteria.

4 . The system of claim 1 , wherein the prompt template comprises a prefix, an attribute value, and a suffix, and triplets are generated from the pseudo-labeled images for training the CSN.

5 . The system of claim 1 , wherein the CSN is further trained using a semi-supervised learning approach with a small percentage of labeled data, the small percentage of labeled data being 1% of the dataset.

6 . The system of claim 1 , wherein the processor is further configured for refining product search capabilities on e-commerce platforms based on user-defined attributes for retrieval of products that align with specific user preferences and requirements using the conditional similarity model.

7 . The system of claim 1 , wherein the processor is further configured for:

identifying and matching one or more tattoos of interest using the conditional similarity model within law enforcement systems for tattoo recognition;

real-time searching through comprehensive law enforcement tattoo datasets to identify associations with gangs or criminal activities based on user-defined notions of similarity, beyond mere visual resemblance; and

recognizing tattoos that share particular attributes but differ in appearance from the one or more tattoos of interest by matching one or more of the tattoos of interest to one or more tattoos from the tattoo datasets based on customizable attributes specified by end users, the attributes being related to specific symbols or meanings.

8 . A computer-implemented method for matching one or more images using conditional similarity pseudo-labels, comprising:

analyzing an unlabeled dataset of images;

accessing a foundational vision-language model trained on a plurality of image-text pairs;

defining a set of attributes each including multiple possible values for generating pseudo-labels based on notions of similarity (NoS);

generating text prompts for each attribute value using a prompt template and encoding the text prompts using a text encoder of the foundational model;

processing each image in unlabeled the dataset of images through a vision encoder of the foundational model to obtain visual features;

comparing the visual features against encoded text prompts to assign a pseudo-label for each attribute for each image; and

training a Conditional Similarity Network (CSN) with the pseudo-labeled images to generate a conditional similarity model.

9 . The method of claim 8 , wherein the foundational vision-language model is Contrastive Language-Image Pre-Training (CLIP).

10 . The method of claim 8 , wherein the set of attributes includes specific user-defined conditions and criteria.

11 . The method of claim 8 , wherein the prompt template comprises a prefix, an attribute value, and a suffix, and triplets are generated from the pseudo-labeled images for training the CSN.

12 . The method of claim 8 , wherein the CSN is further trained using a semi-supervised learning approach with a small percentage of labeled data, the small percentage of labeled data being 1% of the dataset.

13 . The method of claim 8 , further comprising refining product search capabilities on e-commerce platforms based on user-defined attributes for retrieval of products that align with specific user preferences and requirements using the conditional similarity model.

14 . The method of claim 8 , further comprising:

identifying and matching one or more tattoos of interest using the conditional similarity model within law enforcement systems for tattoo recognition;

real-time searching through comprehensive law enforcement tattoo datasets to identify associations with gangs or criminal activities based on user-defined notions of similarity, beyond mere visual resemblance; and

recognizing tattoos that share particular attributes but differ in appearance from the one or more tattoos of interest by matching one or more of the tattoos of interest to one or more tattoos from the tattoo datasets based on customizable attributes specified by end users, the attributes being related to specific symbols or meanings.

15 . A non-transitory computer-readable storage medium comprising a computer-readable program for matching one or more images using conditional similarity pseudo-labels, wherein the computer-readable medium when executed on a computer causes the computer to perform a method, comprising:

analyzing an unlabeled dataset of images;

accessing a foundational vision-language model trained on a plurality of image-text pairs;

defining a set of attributes each including multiple possible values for generating pseudo-labels based on notions of similarity (NoS);

generating text prompts for each attribute value using a prompt template and encoding the text prompts using a text encoder of the foundational model;

processing each image in the unlabeled dataset of images through a vision encoder of the foundational model to obtain visual features;

comparing the visual features against encoded text prompts to assign a pseudo-label for each attribute for each image; and

training a Conditional Similarity Network (CSN) with the pseudo-labeled images to generate a conditional similarity model.

16 . The non-transitory computer-readable medium of claim 15 , wherein the foundational vision-language model is Contrastive Language-Image Pre-Training (CLIP).

17 . The non-transitory computer-readable medium of claim 15 , wherein the set of attributes includes specific user-defined conditions and criteria.

18 . The non-transitory computer-readable medium of claim 15 , wherein the prompt template comprises a prefix, an attribute value, and a suffix, and triplets are generated from the pseudo-labeled images for training the CSN.

19 . The non-transitory computer-readable medium of claim 15 , further comprising training the CSN using a semi-supervised learning approach with a small percentage of labeled data, the small percentage of labeled data being 1% of the dataset.

20 . The non-transitory computer-readable medium of claim 15 , further comprising:

identifying and matching one or more tattoos of interest using the conditional similarity model within law enforcement systems for tattoo recognition;

real-time searching through comprehensive law enforcement tattoo datasets to identify associations with gangs or criminal activities based on user-defined notions of similarity, beyond mere visual resemblance; and

recognizing tattoos that share particular attributes but differ in appearance from the one or more tattoos of interest by matching one or more of the tattoos of interest to one or more tattoos from the tattoo datasets based on customizable attributes specified by end users, the attributes being related to specific symbols or meanings.