Smart Spectrometer

Train an on-device neural network to detect spectral signatures in visible light

To use the app, take a picture. Once the picture is loaded into the app, click on an area of the picture to obtain a spectral sample, which will be taken from the area defined by the 10x10 green box. To train the neural network, enter a label and press "learn." As training progresses, periodicaly press "recognize" to see if more samples need to be entered. Once the network has been trained, the program will either return up to the three closest samples OR indicate that no matches have been found. After sufficient training, the app can be used as an aid in the identification of minerals and chemical samples.

Background information: the app uses an RGB-to-hue algorithm to sample 100 pixels at a time. The spectrograph indicates whether or not the given wavelengths of light have been detected in the sample region. The neural network is based on one-shot learning and cascading typologies and is composed of 3 layers of binary neurons with max-n activations. As the user inputs more samples, the network grows to adapt to them.

Future updates are planned to allow the purchase of ready-made databases for the recognition of different minerals.