"Few-shot learning techniques for challenging applications: automatic mineral and vehicle recognition" 

Andrés Bell

E.T.S. Ing. Telecomunicación, Universidad Politécnica de Madrid, March 2022, "Cum Laude".

Ph.D. Thesis Advisor: Carlos Roberto del Blanco.

In this Ph.D. thesis, few-shot learning techniques have been proposed for object detection and semantic segmentation in innovative high-impact applications in the fields of geology and traffic safety. Their common problem, as in many realistic fields, is the difficulty to elaborate and annotate large databases, having available instead a limited number of samples. This, and given that the growth of few-shot learning encourages a fast and inexpensive deployment of machine learning systems, has led to develop automatic systems for two challenging applications: mineral recognition in drilled cores and in geological samples of minerals or rocks, and nighttime vehicle detection in images acquired by video cameras from traffic surveillance networks. For mineral recognition, two systems have been proposed as part of Innolog, a European Research Project. One of them is focused on processing hyperspectral imagery of drill-core boxes with a machine learning-compatible database creation procedure and an adapted deep neural network. The other system processes infrared and Raman spectral signatures using Siamese Networks and data transformation methods. Regarding vehicle detection, the developed system has been focused on the challenging and critical nighttime scenario. For this purpose, a novel framework based on a grid of foveal classifiers has been designed. Every classifier from the grid processes a global image descriptor (one computed per image) to locate vehicles. Only point-based annotations are required to train the classifiers, speeding up the database creation. Experimental results prove the effectiveness and real-time operation of the proposed systems.

Tweet