Deep Learning

Deep learning is a rather new method of machine learning, that is due to its specifics and excellent performance in image processing often considered as a separate field within artificial intelligence. It works best with big datasets which have multi-dimensional characteristics, and there are a number of such datasets processed at BioSense.

The first group of datasets are hyperspectral, multispectral and thermal images acquired in the lab or during the outdoor field experiments. Here, the the goal was to segment the images and recognise characteristic patterns within them, such as bruises and scratches on apples, infected tomato sepals or ears in images of wheat. Different architectures were tested with various parameters, which included the number of convolutional and fully connected layers, type of activation functions and position of dropout layers. They all were optimised for a particular problem, concerning the size of input images and the number of samples in the database.

The second group consists of airborne particle measurements acquired with Rapid-E, the first automated particle detector based on laser-induced fluorescence technology, which records scattered light and laser-induced fluorescence patterns, representing the shape and chemical footprint of the particles. This is one of the examples of deep neural networks (DNNs) that relied heavily on convolutional layers, due to their ability to deal with datasets fused from different data sources.

In both applications, preprocessing is perhaps the most vital task, as the noisy and non-standardised inputs are very difficult to analyse. Also, an important thing to note is that every problem requires a customised solution, whose complexity corresponds to the richness of the dataset and whose accuracy justifies its learning and execution times.