The Use of Adjacent Video Frames to Increase Convolutional Neural Network Classification Robustness in Stressed Environments
Abstract
The study objective was to use adjacent video frames to increase the robustness of convolutional neural network (CNN) classifiers to stressed targets. We identified and downloaded video clips of targets that moderately change their aspect angle. Video clips of military vehicle target classes were previously used to fine-tune pretrained CNNs through transfer learning. We obtained the frame series from these video clips, and the target in each frame was stressed with different coherent stresses. Instead of relying on the classification of individual frame images, we used different running averages and running products on class probabilities of the classifiers to increase classification robustness to the applied stresses as the aspect angle of the target to the sensor changed. Our results showed modest changes in classifier robustness when we applied moving average/product filters to the output class probabilities. This robustness increase was most pronounced when a small number of elements were averaged, and it regained stability (at an increased robustness) as the filters we applied increased in element size.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 01, 2023
- Accession Number
- AD1205367
Entities
People
- Patrick Debroux