Image recognition allows computers to understand and classify images by analyzing their content. It helps with interpreting, tagging (enrichment with additional metadata), or categorization (sorting into groups) of images.

For example, if you train a dog vs. cat recognition model, and provide it with a folder of mixed pictures, it will be able to distinguish between cats and dogs, and thus sort the images into two separate folders.

It is used in various applications including medical diagnosis, autonomous vehicle systems (object recognition) or security systems (like facial recognition). The accuracy, speed, and efficiency of such systems continue to improve with advancements in artificial intelligence and machine learning techniques and depend on factors such as the quality of training datasets. Deep learning algorithms, particularly convolutional neural networks (CNNs), are commonly used for image categorization tasks.

Ximilar provides several off-the-shelf solutions for classifying specific image data, such as stock photos, home decor and furniture images, fashion photos, or trading and collectible cards. Custom models can be easily trained on our platform. Read the articles in our blog to learn about image recognition technology.

Visual & similarity search technology can analyze the overall visual aesthetic of an image or detected object in an image, independent of the origin of images or metadata (such as keywords). It understands the concept of similarity according to your subjective perception. That is why it can provide the most relevant results to image queries, whether you look for the exact match or similar items.

As for now, the service is able to detect (and mark by bounding boxes) the collectibles such as stamps, coins, banknotes, comic books and trading cards, as well as antique items.

For collectible cards, the service can identify whether it is a Trading Card Game (Pokémon, Magic The Gathering – MTG, Yu-Gi-Oh!, Lorcana, Flesh and Blood and so on) or a Sports Card (Baseball, Basketball, Hockey, Football, Soccer, or MMA), with several additional features (e.g., signature). It can be easily customized to evaluate images based on your criteria. The full taxonomy of the identification results with all supported games and sports can be found in our documentation page.

For comic books, the service can identify more than 1 million magazines, books and manga – via name, title, publisher, issue number and release date.

The service is constantly expanding based on the requests from our customers.

Ximilar’s systems significantly reduce image processing costs by automating repetitive tasks such as analyzing, tagging and sorting of images. This automation results in significant long-term savings, allowing for continuous 24/7 addition of new visual content without additional metadata.

We’re continually enhancing our platform, which enables us to both build services efficiently and quickly and also to modify existing solutions to suit your needs. We reduce costs and labour by utilizing a combination of pre-trained and new models.

Service use is billed via API credits, with a customizable monthly plan based on your consumption. Using our credit calculator helps you optimize cost-effectiveness. For sudden system loads, you can add extra credit packs to your monthly credit supply.

Once your solution is live, we can continually upgrade and enhance it, altering any component in the modular structure. Our feedback mechanisms help us understand which model performs best, allowing us to refine the solution. Also, our ready-to-use solutions are routinely updated to stay industry-relevant, and these upgrades come at no extra cost.

Visual search typically involves searching images or products using an image query, including photos from social media and user-generated content, like smartphone photos. Our Search by Photo combines this technology with detecting individual products in images. One such solution is Search Fashion by Photo.

The technology is frequently employed for similarity searches, particularly for product recommendations in e-commerce. It assesses image or detected object features, like colour, edges, or patterns, to suggest the most similar alternatives.

Another typical use is image & product matching. Since the technology can identify duplicates or nearly identical items in various images with different quality, it can help with curating product galleries and eliminating all unnecessary content.

Even though applications in e-commerce are the most common, visual search provides endless possibilities even in the industrial sector, scientific research or security systems.

Ximilar’s AI solutions were built to handle large datasets containing millions of images. Our cloud doesn’t physically store your image collections for visual search. We process each image once during synchronization, extract descriptive data, and then immediately discard it, allowing for swift image processing. For instance, our largest collections, holding over 100 million images, require only a few hundred milliseconds to search for visually similar items.

Complex images containing more objects, patterns, topics, people or items, can be easily processed with our AI. For instance, our Fashion Search and Home Decor Search solutions automatically detect all fashion apparel or homeware in the images, providing categories and similar stock to each object individually. You can test these solutions in our public demo.

Custom image recognition and detection solutions rely on dedicated training image datasets, whose size depends on the complexity of the problem. For instance, training a model for cat and dog recognition may only require tens of images, while larger projects like our fashion recognition, with thousands of features and tags, necessitate datasets containing thousands of images.

We also enable asynchronous requests. Read more about them in the API Documentation.

Ximilar prioritizes data security and confidentiality. Our customers particularly value the following rules:

• We don’t store user images, except for mutually agreed training datasets, secured on Amazon S3 with time-limited links.
• The intellectual rights to generated image models are shared between you and Ximilar, and we are not authorized to use the models for other customers or internal purposes.
• We adhere to European regulations, including GDPR, and our data center in Prague is multi ISO-certified.
• We can sign NDAs and customize image access restrictions based on your needs.
• Administrators access your image data only when necessary.
• On-premises installation is possible for larger projects.

Feel free to reach out with questions about ethical considerations regarding privacy and data protection.

You can send images in these supported formats to our API (in fields of _base64 or _url): jpg, jpeg, png, webp, heic, bmp, tiff, and jfif. If you upload a gif, only the first frame will be processed. Contact us for customization.

The speed typically ranges from 5 to 100 milliseconds per image for basic image classification and detection models. This timeframe varies based on factors such as input resolution, connectivity, and the speed of the image CDN used. Our platform’s models are generally very fast, and they can be fine-tuned for even greater efficiency on specialized hardware like GPUs and smartphones.