Deciding to identify "similar product images"?
I want to create a cloud solution where I would give an image pool; and then ask "find a similar image for a specific image from this image pool" !! A pool of images can look like "all t-shirts". Hence, similar images mean "a T-shirt with a similar design / color / sleeves", etc.
The label solution will not work as they are at a very high level.
AWS Rekognition gives "face likeness" .. but not "product likeness" .. it does not work like it does for dress images.
I am open to using any cloud provider; but all provide image "tags" which won't help me.
One solution might be that I use some ML framework like MXNet / Tensorflow, create my own models, train them and then use them. But is there any other out-of-the-box solution for any of the cloud providers?
source to share
ibm-bluemix has an api for finding similar images https://www.ibm.com/watson/developercloud/visual-recognition/api/v3/#find_similar
source to share
Using the more intuitive Azure Cognitive Services ( Computer Vision , you can get categories, tags, caption and even more information for images. Processing all of your images will provide tags for your image pool. And this allows you to get similar images based on (multiple) identical tags).
This function returns information about the visual content found in the image. Use domain-specific tags, descriptions, and models to identify and reassure your content. Apply adult / racy settings to enable automatic limitation of adult content. Determine the types of images and color schemes in the pictures.
Example (part) of the result of the Computer Vision API:
Description{
"Tags": [
"train",
"platform",
"station",
"building",
"indoor",
"subway",
"track",
"walking",
"waiting",
"pulling",
"board",
"people",
"man",
"luggage",
"standing",
"holding",
"large",
"woman",
"yellow",
"suitcase"
],
"Captions": [
{
"Text": "people waiting at a train station",
"Confidence": 0.8331026
}
]
}
Tags[
{
"Name": "train",
"Confidence": 0.9975446
},
{
"Name": "platform",
"Confidence": 0.995543063
},
{
"Name": "station",
"Confidence": 0.9798007
},
{
"Name": "indoor",
"Confidence": 0.927719653
},
{
"Name": "subway",
"Confidence": 0.838939846
},
{
"Name": "pulling",
"Confidence": 0.431715637
}
]
source to share
You can also use the Bing Image Search API ( https://azure.microsoft.com/en-us/services/cognitive-services/bing-image-search-api/ ), which allows you to search for images based on specific criteria within Your choice...
source to share
You can use a combination of things here. Use aws rekognition image tag service or any of the above), then create some training data with similar images and load it into something like aws machine learning . This is a bit like what was shown earlier, however I am trying to figure out that although the overlap may not be final, it will be your decision, it will most likely play a role as a transition to a more complex process.
source to share
check the site http://cloudsight.ai/api and try the demo. the sampling result will be
{
"token": "BBKA0lW9O-B2eamXUysdXA",
"url": "http://assets.cloudsight.ai/uploads/image_request/image/314/314978/314978186/79379_86cb4e2611d6b0a3287a926a1ca1fe51_image1_zoom.jpg",
"ttl": 54,
"status": "completed",
"name": "men red and black checkered button-up shirt"
}
{
"token": "bjX7nWGs0toajIDwyvXxlw",
"url": "http://assets.cloudsight.ai/uploads/image_request/image/314/314987/314987168/11.jpg",
"ttl": 54,
"status": "completed",
"name": "blue, gray and navy blue stripe crew-neck T-shirt"
}
source to share