Few-shot logo recognition system
Methods and systems are disclosed for building a few-shot logo recognition system that includes accessing an image with several regions of interest and identifying several objects within the regions of interest using a logo detector neural network. For each object, the logo detector neural network indicates whether the object is a logo. The methods and systems also generate a first and second set of image feature data and a first and second ranked list of logos. A final ranked list of logos is generated based on the first and second ranked list of logos and a category associated with each logo in the final ranked list of logos is identified.
1 . A system for few-shot logo recognition comprising:
one or more processors;
at least one memory component storing instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
accessing an image comprising a plurality of regions of interest;
identifying a plurality of objects within the plurality of regions of interest using a logo detector neural network, wherein for each object of the plurality of objects, the logo detector neural network is trained to generate an indication that the object is a logo;
generating a first set of image feature data for the image and a second set of image feature data for the plurality of regions of interest;
generating a first ranked list of logos from the identified plurality of objects, the first ranked list of logos generated based on matching the first set of image feature data with image feature data associated with a database of logos;
generating a second ranked list of logos from the identified plurality of objects, the second ranked list of logos based on matching the second set of image feature data with the image feature data associated with the database of logos;
generating a final ranked list of logos based on the first ranked list of logos and the second ranked list of logos; and
identifying a category associated with each logo in the final ranked list of logos.
2 . The system of claim 1 , wherein each logo in the final ranked list of logos is associated with a similarity score, the operations further comprising:
identifying a subset of logos in the final ranked list of logos, the subset of logos identified based on having similarity scores outside of a predefined range;
determining that the identified subset of logos are false positive logo objects using a geometric verification algorithm; and
in response to determining the identified subset of logos are false positive logo objects, removing the identified subset of logos from the final ranked list of logos.
3 . The system of claim 2 , wherein the similarity score is a cosine similarity score.
4 . The system of claim 1 , wherein the generating the first set of image feature data for the image further comprises:
generating a plurality of cropped images, the cropped images generated by cropping each region of the plurality of regions from the image;
for each cropped image, providing the cropped image as input to a feature extractor neural network trained to:
generate a plurality of resized images, each image in the plurality of resized images resized at a different scale;
center crop each image in the plurality of resized images;
generate a plurality of normalized resized images from the plurality of resized images; and
generate image feature data for each image in the plurality of normalized resized images; and
generating first multi-scale image feature data for the image by aggregating the image feature data for each image in the plurality of normalized resized images.
5 . The system of claim 1 , wherein the logo detector neural network is further trained to generate a set of pixel coordinates representing a boundary of the object.
6 . The system of claim 1 , wherein the generating the second set of image feature data for the plurality of regions of interest further comprises:
for each region of interest in the plurality of regions of interest:
generating a plurality of cropped images, the cropped images generated by cropping portions of the region of interest;
for each cropped image, providing the cropped image as input to a feature extractor neural network trained to:
generate a plurality of resized images, each image in the plurality of resized images resized at a different scale;
center crop each image in the plurality of resized images;
generate a plurality of normalized resized images from the plurality of resized images; and
generate image feature data for each image in the plurality of normalized resized images; and
generating second multi-scale image feature data for each region of interest by aggregating the image feature data for each image in the plurality of normalized resized images.
7 . The system of claim 1 , wherein the logo comprises at least one of a symbol, word, or name.
8 . The system of claim 1 , wherein the database of logos comprises at least one manually labeled logo image.
9 . The system of claim 1 , wherein the operations further comprise:
accessing a second image from a computer device;
analyzing the second image, the analysis comprising an identification of a category associated with at least one logo identified in the second image; and
in response to analyzing the second image, causing display of an interactive window on a graphical user interface of the computer device, the interactive window comprising an indication of the category associated with the at least one logo identified second image.
10 . A method for few-shot logo recognition comprising:
accessing, by one or more processors, an image comprising a plurality of regions of interest;
identifying, by the one or more processors, a plurality of objects within the plurality of regions of interest using a logo detector neural network, wherein for each object of the plurality of objects, the logo detector neural network is trained to generate an indication that the object is a logo;
generating a first set of image feature data for the image and a second set of image feature data for the plurality of regions of interest;
generating a first ranked list of logos from the identified plurality of objects, the first ranked list of logos generated based on matching the first set of image feature data with image feature data associated with a database of logos;
generating a second ranked list of logos from the identified plurality of objects, the second ranked list of logos based on matching the second set of image feature data with the image feature data associated with the database of logos;
generating a final ranked list of logos based on the first ranked list of logos and the second ranked list of logos; and
identifying a category associated with each logo in the final ranked list of logos.
11 . The method of claim 10 , wherein each logo in the final ranked list of logos is associated with a similarity score, the method further comprising:
identifying a subset of logos in the final ranked list of logos, the subset of logos identified based on having similarity scores outside of a predefined range;
determining that the identified subset of logos are false positive logo objects using a geometric verification algorithm; and
in response to determining the identified subset of logos are false positive logo objects, removing the identified subset of logos from the final ranked list of logos.
12 . The method of claim 11 , wherein the similarity score is a cosine similarity score.
13 . The method of claim 10 , wherein generating the first set of image feature data for the image comprises:
generating a plurality of cropped images, the cropped images generated by cropping each region of the plurality of regions from the image;
for each cropped image, providing the cropped image as input to a feature extractor neural network trained to:
generate a plurality of resized images, each image in the plurality of resized images resized at a different scale;
center crop each image in the plurality of resized images;
generate a plurality of normalized resized images from the plurality of resized images; and
generate image feature data for each image in the plurality of normalized resized images; and
generating first multi-scale image feature data for the image by aggregating the image feature data for each image in the plurality of normalized resized images.
14 . The method of claim 10 , wherein the logo detector neural network is further trained to generate a set of pixel coordinates representing a boundary of the object.
15 . The method of claim 10 , wherein generating the second set of image feature data for the plurality of regions of interest comprises:
for each region of interest in the plurality of regions of interest:
generating a plurality of cropped images, the cropped images generated by cropping portions of the image;
for each cropped image, providing the cropped image as input to a feature extractor neural network trained to:
generate a plurality of resized images, each image in the plurality of resized images resized at a different scale;
center crop each image in the plurality of resized images;
generate a plurality of normalized resized images from the plurality of resized images; and
generate image feature data for each image in the plurality of normalized resized images; and
generating second multi-scale image feature data for each region of interest by aggregating the image feature data for each image in the plurality of normalized resized images.
16 . The method of claim 10 , wherein the logo comprises at least one of a symbol, word, or name.
17 . The method of claim 10 , wherein the database of logos comprises at least one manually labeled logo image.
18 . The method of claim 17 , wherein the database of logos comprises a set of retrieved logo images, the set of retrieved logo images generated by:
providing the at least one manually labeled logo image as a query input to a database of images; and
based on the query input, identifying visually similar images to the query input.
19 . The method of claim 10 , further comprising:
accessing a second image from a computer device;
analyzing the second image, the analysis comprising an identification of a category associated with at least one logo identified in the second image; and
in response to analyzing the second image, causing display of an interactive window on a graphical user interface of the computer device, the interactive window comprising an indication of the category associated with the at least one logo identified second image.
20 . A non-transitory computer-readable storage medium storing instructions that, when executed by at least one processor, cause the at least one processor to perform operations for few-shot logo recognition comprising:
accessing an image comprising a plurality of regions of interest;
identifying a plurality of objects within the plurality of regions of interest using a logo detector neural network, wherein for each object of the plurality of objects, the logo detector neural network is trained to generate an indication that the object is a logo;
generating a first set of image feature data for the image and a second set of image feature data for the plurality of regions of interest;
generating a first ranked list of logos from the identified plurality of objects, the first ranked list of logos generated based on matching the first set of image feature data with image feature data associated with a database of logos;
generating a second ranked list of logos from the identified plurality of objects, the second ranked list of logos based on matching the second set of image feature data with the image feature data associated with the database of logos;
generating a final ranked list of logos based on the first ranked list of logos and the second ranked list of logos; and
identifying a category associated with each logo in the final ranked list of logos.