Image interpretation in Secchi3000
Image interpretation in Secchi3000
Track 02: Artificial intelligence, analytics, and value from data
Tags: water system monitoring, water quality, measurement devices, pattern recognition, image processing, machine vision.
Sponsor: SYKE
Challenge
Analyze the picture datasets with available tools (e.g. IBM Watson). Try to automatically locate the panels/color areas and extract the statistical data of the different color components. Additionally, try to recognize the failed images.
Develop an automatic method to reproduce the previously manually done image interpretation. You can either develop algorithms based on the existing Secchi3000 Matlab code or train an algorithm with the available training datasets.
Background
Secchi3000 data interpretation is based on extracting representative statistically descriptive red, green, and blue pixel color component values from the differently colored areas under even light conditions. A successful, automated process should recognize if the interpretation is not possible.
In the learning datasets, manually processed images are converted into turbidity values by an algorithm that utilizes data from two panels at different water depths and the red-green-blue colour components from the black-grey-white coloured areas.
Secchi3000 images proved difficult to interpret automatically using straightforward image processing methods. However, the task itself is simple: the images contain rectangular color areas on panels, or as in the test dataset, black-gray-white areas in two panels at different depths.
There was significant confusion if the images changed between the landscape and portrait orientation, if the image area was partly shadowed, or if the lower panel disappeared due to too turbid or dark water. Additional confusion was caused by error images not quite covering all the panels or not covering the panels at all.
Additional bonus is available for system setups and solutions that are compatible with Secchi devices with different colour setups and different panel patterns.
Approaches and considerations
Try to find rectangles in the image. Next, try to associate them according to size and apparent color differences. Three similar sized rectangles should be found, and there should be two groups.
Note: The structures and patterns or internal angles/orientations in different devices are not necessarily identical aside of the three-color-area patterns which are standard.
Note: There could be reflections coming from the front window.
Data
Test datasets of iQwtr measurements, including original images, interpreted results, and software for basic manual analysis.
https://drive.google.com/drive/folders/1pb9luGHh_I_pp0XHQfc9w0fKdiUSSxgf?usp=sharing
Additional materials
Detailed description and background for Secchi3000
https://drive.google.com/file/d/11tlFMXlKHZmE0UbvS_Q484NJS8o2sqkE/view?usp=sharing
Results and market reviews of iQwtr implementation of Secchi3000 devices, associated documentation.
https://drive.google.com/drive/folders/1ejrhmxOSww3D9HQqDrGPItCiIlBl1aPC?usp=sharing
Technology
You can consider and try out the below examples and technologies. However, you’re not limited to use them at all, or only them. You should explore different approaches and find technology that supports your idea.
IBM DSX / Machine Learning
http://heidloff.net/article/watson-machine-learning-sample
https://datascience.ibm.com/community
Watson APIs
https://www.ibm.com/watson/developercloud/doc/index.html
https://www.ibm.com/watson/developercloud/starter-kits.html
Watson Visual Recognition
https://github.com/IBM-Bluemix/Visual-Recognition-Tile-Localization
Bluemix service catalog (e.g. Watson and IoT)
https://console.bluemix.net/catalog/
IBM developer docs
https://console.bluemix.net/docs/
https://www.ibm.com/developerworks/
Matlab (trial)
https://se.mathworks.com/programs/trials/trial_request.html?requestedDomain=www.mathworks.com