Our Prize Winning AWS Hackathon Entry

This text is written to clarify my venture that I submitted to the AWS Hackathon on Devpost. Me and my good friend had initial started the mission as a learning step to build and deploy a machine learning undertaking. We ultimately ended up successful the runner up for Computer Vision for our efforts. With speedy urbanization and industrialization Air pollution has develop into an alarming challenge globally. Among totally different air pollutants particles , matter in (PM2.5) measurement range are able traverse deeply into our respiratory tract , transmitting hazardous chemicals into the human lungs and blood inflicting respiratory and cardiovascular well being points . We goal to develop image-based air quality analysis software , specifically a one that estimates the focus of particulate matter with diameters in vary of 2.5 micrometers. It may possibly let you know what your respiratory from images generated using consumer grade hardware. Convolutional Neural Network (CNN) with the high computational power of Amazon EC2 DL1 occasion to resolve a pc imaginative and prescient downside of classifying natural photos into three different categories ;wholesome , moderate and danger based mostly on their PM2.5 concentrations.



We created our personal dataset by combining knowledge from 2 sources . We realized that, some reasons for overfitting could possibly be high imbalance in knowledge , we had 2232 wholesome pictures the place as the moderate and hazard categories had only 1480 and 591 pictures respectively . 2232. Then we trained our model following the identical methodology as before but this time we used ImageNet weights un-freezed the last a hundred layers for BloodVitals SPO2 coaching , then superb-tuned the mannequin with a lower studying fee ,we additionally used early stopping callback with a persistence of 3 . Now, there are some more things we now have accomplished to improve our model. We migrated our mannequin onto the dl1 occasion and initially skilled utilizing 1 gaudi processor. This significantly improved our efficiency and efficiency .We had been able to quickly take a look at , experiment and tweak changes . 76 %. We then converted our mannequin right into a tensorflow lite model and used a huggingface demo tool called gradio to quickly demonstrate our software.



Initially as we tried training our model we observed a number of the image information in the dataset were corrupted .We automated the means of scanning via and deleting these information and then proceeded to coaching model . We used EfficientNet structure since its renown for its efficiency in low spec machines, BloodVitals SPO2 this known as transfer studying . 30 layers and set the weights parameter to false. 60 % however our training accuracy was high, the model could be overfitting. As we have now seen we have been ready an accuracy of about 76 % which is pretty good contemplating the information we had. Now , why do we want this software and the way is it different from the present air monitoring system methods, we already know Exposure to wonderful particles may cause long run well being results reminiscent of lung and heart downside but in addition they cause quick term health effects such as eye, nose, throat irritation.



The existing air quality monitoring methods are extremely dependent on monitoring stations, which are once more situated far away because of the excessive setup price and expensive equipment. For a particular location, the air high quality relies on how air moves by means of the area and the way persons are influencing the air , so we can not depend on something that was predicted some time back or in some place that isn't our site of interest .This utility helps you analyse the air high quality around you with a single picture and more importantly is portable. We are able to carry it anyplace we want to and know what your being exposed to . Honestly I don’t know, I think there is potential for this app to combine into our routine. People may want features to help plan their day out, integrate with their phone’s calender and suggest the best time to plan an outside exercise. Right now the Artificial intelligence that runs the machine runs on the azure cloud. In the future we would like to be able to carry it into the app and run it natively.



Certain constituents within the blood affect the absorption of gentle at numerous wavelengths by the blood. Oxyhemoglobin absorbs light more strongly within the infrared area than within the purple region, whereas hemoglobin exhibits the reverse habits. Therefore, highly oxygenated blood with a excessive concentration of oxyhemoglobin and a low focus of hemoglobin will tend to have a high ratio of optical transmissivity in the pink region to optical transmissivity within the infrared region. These alternating portions are amplified and then segregated by sampling devices working in synchronism with the red/infrared switching, so as to offer separate indicators on separate channels representing the red and infrared mild transmission of the physique structure. After low-move filtering to remove signal components at or above the switching frequency, every of the separate alerts represents a plot of optical transmissivity of the physique structure at a particular wavelength versus time. AC part brought on only by optical absorption by the blood and varying on the pulse frequency or coronary heart price of the organism.