Processing of images from various sources containing different angles of the designs, varying resolution, and context to conclude in an automated manner into evaluation results aligned with the agency’s business unit requirements.
In order to mimic the processes of the agency’s business unit, we broke down their thought process based on workshops and data science reports. Then we searched the literature for open-source implementations of each phase/ component required and we synthesized it in an end-to-end solution. Since this is a vision-based analytical process, pre-trained convolutional neural networks were the technologies of choice for most of the tasks accompanied with image processing techniques.
We split the problem into a pre-processing and a post processing phase. Initially, each image passed through a segmentation phase where pre-trained CNN model was used to automatically detect its boundaries. An additional layer of pre-processing was then used to identify and remove objects that are “neutral”, meaning they are expected to be found in the background of such design. After this, the clean part of the original image is split into different modeling approaches that generated a “score” for each of the rules that were provided by the business unit. Configuration of each score threshold can be used to control the “sensitivity” of the model and fine-tune based on the needs of the system. The solution is used as a python microservice and is deployable via Kubernetes or other docker-based installation. This way, it can be easily integrated in the agency’s front-end or back-end application providing assistance to the site’s user as well as the business unit.
Deficient Design Images Detection Model
DESCRIPTION WHAT WE HAVE DONE/ACHIEVED
We have successfully translated business requirements into system components based on pre-trained AI models and image processing. We have performed field tests of the solution with resulting accuracy reaching up to 73% for hundreds of cases including thousands of images. We have successfully deployed our solution in the agency’s production environment.