Image Analysis with Small Data
The objective of this research project is the development of a detection algorithm using thermal images, which have the advantage of preserving the privacy of individuals and allow operation even at night.
The Challenge
Training a neural network would require a large sample of annotated thermal images. Unfortunately, the availability of such datasets is extremely limited, especially when compared to the amount of data available for networks that, in contrast, use RGB images.
The Solution
We exploit domain adaptation techniques to adapt neural networks to operate on new domains. In this case we start from networks trained on RGB images and modify the first layers through a bottom-up process of convolutional network adaptation. In this way the new model is able to operate detection in the thermal spectrum. Furthermore we use a task-conditioned approach to automatically recognize if the image under examination is day or night allowing the algorithm to adapt to the specific task.
Results and Benefits
Our approach allows us to overcome the limitations imposed by the lack of labeled data. Transferring knowledge from a network trained on RGB data to the thermal domain, allows us to achieve performances similar to those achieved with the use of RGB images but with the advantage of preserving privacy and the possibility of use in night conditions.
