Underwater imagery is used in many fields like marine biology, ecology, oil & gas, and coastal planning to name a few. These fields all have unique data types, requirements and expected data products. ViQi offers solutions for projects with different requirements through our platform. ViQi is generic enough to support many workflows but is also customizable to make each project feel specialized.
Percent Cover Automation
One of the most commonly used quantification techniques used in underwater image analysis is percentage cover quantification. ViQi's platform offers tools to easily manage, find and annotate unlimited deployments. Our web based platform is available anywhere with internet and a web browser and allows multiple users to work concurrently. This speeds-up the annotation process required to train and validate AI solutions.
The following screen shot demonstrates the web-based annotation system that supports any number of arbitrarily large images and any class naming and hierarchy.
While datasets are being manually annotated, they can be immediately used for data products such as species occurrence histograms, proliferation maps and many others.
Continuous Annotation & Validation
The graph below shows more than 300 different taxonomic types manually identified in the dataset of 2.4K images.
Some classes may contain many (20K) identified samples while the other ones may have fewer than 3. While a very large number of training samples is always helpful it might be unnecessary for a particular task. On the other hand 3 samples is not nearly enough to train and validate an AI system.
ViQi's approach allows users to continuously annotate and validate the automated model to identify what classes need more annotations.
At the same time, models trained at different sets can be kept and compared. This way automated classification can be versioned and the same datasets can be compared with different versions for validation and evolution.
AI Models In Just A Few Clicks
The initial training dataset can be easily created by selecting data from specific deployments of interest and distributing those images for annotators. When some number of annotations of taxonomic classes of interest have been created an AI model can be trained and validated with just a few clicks in the user interface.
Once training and validation are finished it is easy to verify the identification performance, see classes creating confusion and increasing errors and ultimately ignore classes that perform poorly. This way it's quick and easy to create a model that can perform annotation work at a high degree of confidence.
Due to uneven distribution of species abundance, a model that reliably identifies 15 taxonomic classes may be able to produce 70% of all the annotation work while saving a lot of time, money, and most importantly being reliable and reproducible.
Once a model is created and validated it can be run at a large scale, creating data products quickly. It can even automatically annotate all uploaded images of a specific site. This way data products can be available the next morning after a survey data was uploaded completely automatically.
Moreover, ViQi can score every automated sample and thus by discarding uncertain samples can dramatically increase performance if more samples can be obtained.
For the shallow benthic imagery example above starting with 340 available taxonomic classes. After training classes with at least 100 annotated samples, 117 in total, the average accuracy was at 59% and error of 23%.
By ignoring classes performing poorly and accepting the top 63 classes the model performed at 69% accuracy and 19% error.
To further improve annotation quality we can discard samples below 90% confidence achieving 97% accuracy and 4% error.
Here is an example of a very dense percent coverage where colors mean different taxonomic classes. When discarding samples to increase accuracy, percent cover points may simply be left not annotated.
Segmentation And Other Data Products
The trained AI model can now perform many tasks, not just percent cover classification as it was trained. Additional data products include partitioning and segmentation.
Let's look at another dataset to better understand system strengths and weaknesses and how we can go around them. We will use a dataset comprised of approximately 5K low-resolution and low-contrast deep benthos images with >70,000 annotations for object detection and identification as well as substrate classification and partitioning.
Two models were created for this case, one to identify types of substrate, like sand, pebble, rocks, etc . . . The other model was trained on objects of interest like species of fish, brittle stars, etc. The object identification model demonstrated average accuracy of 26% and error of 21% for 56 classes (with at least 100 samples) out of 170 available classes in the annotations. The lower performance in this case is mostly due to low image quality and resolution.
Although, the top 5 most performant classes in this model demonstrate accuracy of 90% and error of 5%. The Connoisseur system allows discarding classes that are not performing well in validation substantially increasing overall accuracy. This way the model will be producing fewer annotations but they will be a lot more robust requiring less time to verify automated outputs and saving time with manual annotation.
In the following examples it's possible to see how low resolution is affecting detection closer to image borders due to the lack of pixels, how brittle stars are combined into large spanning regions and rough substrate partitioning.
ViQi's Connoisseur system is robust and flexible and can be tuned to a variety of problems reliably automating manual work. The amount of automation will greatly depend on quality and amount of input and training data. ViQi is always happy to help you achieve quality results for your solution, contact us today to get a demo.
Learn more about ViQi Connoisseur for your machine learning needs.