This Will Be Our Next-Gen Visual Search Engine- Part 1

Photo by JOSHUA COLEMAN on Unsplash

We've been working on this for a long time - and we still do. The next major release of our visual search engine: pixolution flow 4. It's widely used for asset discovery and retrieval as well as de-duplication of image collections. As we're starting to get into the final phase it's time to reveal some details about what we have built and what you can look forward to.

Part 1

Since pixolution flow 4 will be packed with quite a lot of innovations we split the release preview into two parts. This first part focuses on new features and capabilities made possible by a complete redesign of pixolution flow.

Modular Architecture

pixolution flow 4 is rewritten from scratch and contains all our lessons learned over the past years. The most important change is a modular architecture providing great flexibility and customization opportunities for future developments.

Modules adding functionality to pixolution flow

Essentially, pixolution flow becomes a hub for modules which encapsulate functionality. For example visual image search, multi color search or text space filter are capsuled in modules. Each module contains the algorithm to process its supported input (e.g. images, text or videos), provides search or filter logic, how to store its output in the Solr index and extends the API of pixolution flow to expose its features to the user.

Doing so has several advantages for you: Updating pixolution flow will be easier. When we have a better AI descriptor and deliver a new version of the visual image search module, you can plug it to Solr and run two different module versions of the same functionality side by side. You can lazily analyze your images with the new module while using the current one for your live search. Whenever you are ready to switch to the newer version you just have to change a module version parameter in your search query.

We support new Solr versions faster, since our Solr specific code layer is reduced to a minimum to make upgrades easier for us.

But the greatest impact of a module based architecture is the possibility to build and integrate customized AI modules for you. We can train individual AI models adapted to your individual data set and needs and extend your pixolution flow instance with it. Let's see below what these modules might be.

Custom AI Modules

Whether to solve classification problems or provide new search capabilities, the possibilities with custom AI modules are endless. To give you a feeling of what we could train and integrate in a custom module, here are some examples.

Recognize certain objects and logos in images to get product recommendations, identify the concrete product based on photos or check occurence of your brand. This can be applied to all companies handling specific image collections like fashion shops or industry products.

Identify the used font type in an image containing text to automatically check corporate design guidelines or clear rights for font usage. Check out how we already built such a font recognition engine.

Detect persons in images and identify images which might need special attention when storing or using them to comply with data protection regulations like GDPR.

Classify image content based on individual categories. This might be automatic categorization of uploaded hotel images or real estates (e.g. bedroom, bathroom, reception, balkony etc.) for better presentation and search in online portals or to check whether a dataset contains all required photos.

These are only some examples and we are looking forward to hear from your exciting challenges you want to solve.

Multi-threaded analysis

Solr has the restriction to analyze and index documents sequentially. This is fine when we deal with text documents, since they don't require much computation power to process. However, analyzing images or videos is much more complex and therefore time consuming. pixolution flow processes incoming images several times (e.g. each module analyzes the image), and with our upcoming release it will be possible to parallelize the analysis on a document level. This leads to faster indexing.

Stay tuned for part 2

We hope we could give you a good first overview about pixolution flow 4. But the best is yet to come. In the second part of our release preview we will show you how our search features will massively improve. Stay tuned for part 2!


Get our latest blog posts and updates straight to your inbox.