Visual quality performance predictors

1 . One or more computer storage media storing computer-readable instructions that when executed by a processor, cause the processor to perform operations for a visual search recommendation, the operations comprising:

receiving a search image comprising an item associated with a first aspect as a search query;

providing the search image as an input into an image quality indication model, wherein the image quality indication model is trained on embeddings generated from a plurality of images comprising items, wherein the plurality of images are labeled based on image quality aspects comprising at least one of blurriness, background quality, watermarks, angle of view, or inclusion of human body parts, to generate an output of an image quality indication for the item within the search image, the item defining a portion of the search image, the image quality indication generated from the first aspect corresponding to the portion of the image defined by the item;

identifying a subset of images comprising the item and not comprising the first aspect from an image corpus that have the item, wherein the item in each image of the subset of images has a corresponding image quality indication above a threshold;

determining a search query performance for the search image based on the image quality indication of the first aspect of the item within the search image and based on an image similarity between the search image and each of the subset of images;

identifying a first image from the subset of images that exceeds the search query performance; and

causing to display the first image as a recommendation for the search query.

2 . The media of claim 1 , the operations further comprising:

ranking the search query performance for each of the subset of images;

based on the ranking, causing to display a plurality of images, from the subset of images that each exceed the search query performance, as recommendations for the search query;

receiving a selection of at least one of the recommendations for the search query; and

causing to display search results based on using the at least one of the recommendations as the search query.

3 . The media of claim 1 , further comprising:

retrieving a set of search results from the image corpus based on receiving a selection of the first image as the recommendation for the search query;

determining a post-retrieval image similarity for each of the set of search results based on a comparison between the first image and images for the set of search results; and

causing to display a portion of the set of search results based on the post-retrieval image similarity for each of the set of search results.

4 . The media of claim 1 , further comprising:

identifying the first aspect of the search image and the first aspect within the subset of images from the image corpus using a deep neural network to detect an object having the first aspect within the search image and the subset of images.

5 . The media of claim 4 , wherein the search image comprises a second aspect, wherein the subset of images is identified using the deep neural network to detect the object having the first aspect and the second aspect, and wherein the operations further comprise:

determining that the first image exceeds the search query performance based on:

an image quality indication associated with the first aspect and the second aspect, provided by the image quality indication model, for the first image; and

an image similarity between the search image having the first aspect and the second aspect and the first image having the first aspect and the second aspect.

6 . The media of claim 1 , the operations further comprising:

determining the image similarity for the search query performance by applying an image similarity determination model, including a convolutional neural network, to the search image and each of the subset of images for comparing one or more categories associated with the search image and each of the subset of images; and

determining the search query performance for the search image based on the image similarity determined by applying the image similarity determination model.

7 . The media of claim 1 , wherein the image quality indication corresponds to how fast a search may be performed using a corresponding image.

8 . The media of claim 1 , wherein the search query performance corresponds to a response-time and error rate associated with providing search results for the search query.

9 . A computer implemented method for a visual search recommendation, the method comprising:

receiving a search image comprising an item associated with a first aspect for a search query;

providing the search image as an input into an image quality indication model, wherein the image quality indication model is trained on embeddings generated from a plurality of images comprising items, wherein the plurality of images are labeled based on image quality aspects comprising at least one of blurriness, background quality, watermarks, angle of view, or inclusion of human body parts, to generate an output of an image quality indication for the item within the search image, the item defining a portion of the search image, the image quality indication generated from the first aspect corresponding to the portion of the image defined by the item;

identifying images from an image corpus that comprise the item and do not comprise the first aspect, and have an image quality indication for the item within the search image, the item defining a portion of the search image, wherein the identified images have an image quality indication above a first threshold and that have an image similarity with the item within the search image that is above a second threshold;

determining that a first image from the images from the image corpus exceeds a search query performance based on the image quality indication and the image similarity for each of the images; and

providing the first image as a recommended search image for running the search query.

10 . The method of claim 9 , wherein determining the image quality indication further comprises detecting motion blur using an image edge detector that detects a hue angle difference between two hue angles corresponding to the search image.

11 . The method of claim 9 , wherein determining the image quality indication further comprises determining a background quality using quantitative characteristics including at least one of gradients of brightness, gradients of color, gradients of saturation, or values obtained by applying a set of digital filters to pixels of a background portion.

12 . The method of claim 9 , further comprising:

determining the image similarity for each of the images from the image corpus by applying an image similarity determination model, including a convolutional neural network, to compare one or more categories associated with the search image and each of the images; and

determining that the first image exceeds the search query performance based on applying the image similarity determination model.

13 . The method of claim 9 , wherein determining the image quality indication further comprises using user usage metadata associated with the search image including at least one of: number of times viewed, number of edits, number of transmissions, or total viewing time since capture or upload.

14 . The method of claim 9 , wherein identifying the images from the image corpus further comprises determining a relative size of the item within the images.

15 . A system for a visual search recommendation, the system comprising:

at least one processor; and

one or more computer storage media storing computer-readable instructions that when executed by the at least one processor, cause the at least one processor to perform operations comprising:

receiving a search image as a search query;

prior to executing the search query, determining, using an image quality indication model, an image quality indication for the search image, wherein the image quality indication model is trained on embeddings generated from a plurality of images comprising items, wherein the plurality of images are labeled based on image quality aspects comprising at least one of blurriness, background quality, watermarks, angle of view, or inclusion of human body parts;

prior to executing the search query, identifying images, comprising the item and not comprising the first aspect, from an image corpus that have a corresponding image quality indication above a threshold, wherein the corresponding image quality indication is for the item within the search image, the item defining a portion of the search image;

prior to executing the search query, determining a search query performance for the search image based on the image quality indication of the search image and an image similarity between the search image and each of images having the corresponding image quality indication above the threshold;

prior to executing the search query, identifying a first image from the images that exceeds the search query performance; and

prior to executing the search query, providing the first image as a recommendation for the search query.

16 . The system of claim 15 , the operations further comprising:

prior to executing the search query replacing the search image with the first image, and executing the search query based on the first image.

17 . The system of claim 15 , the operations further comprising:

identifying search results based on receiving a selection of the first image provided as the recommendation for the search query;

determining a post-retrieval image similarity for each of the search results based on a comparison between the first image and images of the search results; and

providing a portion of the search results based on the post-retrieval image similarity for each of the search results.

18 . The system of claim 15 , the operations further comprising:

prior to executing the search query, identifying a first aspect and a second aspect of the item within the search image, wherein the first aspect and the second aspect correspond to a portion of the image defined by the item;

prior to executing the search query, identifying the images from the image corpus based on each of the images having the first aspect and the second aspect; and

prior to executing the search query, determining the search query performance for the first image based on the corresponding image quality indication for the first image and the image similarity between the first aspect and the second aspect of the search image with the first aspect and the second aspect of the first image.

19 . The system of claim 15 , the operations further comprising:

prior to executing the search query, determining the search query performance for the first image based on the corresponding image quality indication for the first image and the image similarity between one or more category embeddings of the first image with the search image.

20 . The system of claim 15 , wherein the images from the image corpus are identified based on the images having a category embedding associated with a category of the search image.