Original text in Spanish (from the ReadMe of my project hosted on GitHub) and automatically translated into English using Google Translator:
Purpose:
Jordi Garcia Castillón Project - Inclusion and accessibility (Hackathon for Good AWS - Intel): Computer vision for people with prosopagnosia and for other diseases and/or disabilities.
Proposed solution and social impact:
This system is simply intended to be a base application to "provide" any other system or element (as mentioned in the previous points) with the functionality for which it has been prepared, and all for the following:
Although the application of this system can also be used for the assumption of other diseases (such as, for example, Alzheimer's where a person does not remember a person if he sees it and the system, if voice were incorporated, could recognize the image and tell him who it is ) or other diseases or disabilities, primarily this application has been proposed for people who suffer from the disease called prosopagnosia.
Prosopagnosia is a rare disease that can be both congenital (it is estimated that it affects 2.5% of the population) and acquired (due to cerebral vascular accidents, tumors, Parkinson's, etc.), for which the person she becomes unable to facially recognize the person in front of her (or even herself in a mirror).
This difficulty in recognizing another person, in addition to implying a clear reduction in the quality of life for the affected person, imposing strong limitations and having a high negative impact on a psychological level, can endanger their own physical integrity. For example, a person with prosopagnosia who is at home and cannot recognize if when he opens the door that person is known or not or if, even, it may become someone who wants to harm him in some way (someone who has been separated sentimentally due to abuse, violence or many other cases or assumptions).
The solution presented here through an easy recognition system allows the affected person to establish categories of people so that when they knock on the door or speak to someone on the street (to give just two examples) they can know automatically if the person subject in front is a family member, an acquaintance or if he is a person in danger and must move away or request help and assistance. In short, it alerts him and puts him on notice of who she may be and prepares him to act accordingly.
Evidently, this system has only been developed in a first incipient phase and additional layers could be incorporated that would allow the person to know in even more detail that person who, due to her illness, is unable to know by herself.
The additional layers mentioned above could include, for example, that the system directly recognize which person it is (his father, mother, friend's name, etc.), or establish a probability that it was part of his closest environment ( for example, by detecting patterns in traits between people who have a relationship of kinship and consanguinity between them, among others).
For all this, we have worked on a primary database of images that have been labeled for each of the desired groups and a data augmentation technique has been applied (with Roboflow) both to increase the base and to make the algorithm more "resistant". before possible changes, distortions, variants that may exist in the camera or in the subject to be recognized. In the repository you will find all the information about it (some of the json files have been modified to eliminate potential identifying data of the resource).
Technical introduction and disclaimer:
This project has been developed solely for prototype demonstration purposes for the Hackathon for Good, organized by AWS and Intel in March 2022.
The primary open data source is the well-known LFW image dataset from the University of Massachusetts.
The current state of development does not enable it at this time for its medical uses, orientation or any health function.
Both the results obtained and the functionality of this development can only be interpreted for now as a mere basic demonstration, since it is in a functional state but obviously neither the number of images with which the system has been prepared nor many other variables or Considerations mean that the production reliability of this development is now far from established.
Required Sections:
- What is the chosen challenge and the value of the solution in its resolution?
The category submitted to the contest is “Inclusion and accessibility”.
The value that the solution is considered to provide is that it manages to provide a response that can become of vital importance and provide, in any case, an improvement in the quality of life of the people who They suffer from a disease that prevents them from recognizing people.
- What services work with AWS and Intel and why were they chosen?
Essentially, we have worked with AWS Rekognition and this modality has been chosen because it is a service managed and completely managed by AWS that allows us to prepare a model quickly, generating in very little time a fully functional and operational application that can later be easily reproducible, scalable and accessible in many ways.
Using AWS Rekognition for image recognition allows you to enable the model (for example through its API) to a multitude of systems that you can think of, from WebApps to cameras and systems that the sick person can use for their use.
Obviously, all or almost all systems can run on Intel processors, but the great value of Intel in this solution is not only in this capacity, but also in the possibility of taking the model to specific Intel solutions such as SBC boards that run on Intel in general and Intel Coffee Lake in particular, and even more in proprietary solutions to which the solution can be implemented in its Edge modalities with Intel NUC minicomputers and any Webcam (this can be any Intel if desired).
Final Notes:
The members of the jury can request access to the resource in Rekognition, to the API at the time they wish to be able to check the operation of the system if they wish, as established in the rules of the contest.
Link to the video of the project: Youtube
Top comments (0)