Invited Seminar — Oxford Mathematical Institute (2025)

I delivered an invited seminar in the Machine Learning and Data Science Seminar Series at the Oxford Mathematical Institute, presenting my work on bridging classical and modern computer vision for tree crown semantic segmentation.

The talk covered:

• Real-world complex scenes challenges (lighting variation, occlusions, spectral ambiguity, scale differences).
• Deep learning feature bias in CNNs and its impact on generalisation.
• Comparison of CNNs, Gabor filters, and Log-Gabor filters.
• Theoretical limitations of classical Gabor filters.
• Introduction of LogGabConv, a trainable Log-Gabor parameterised convolutional layer.
• A new backbone architecture (PerceptiveNet) with averaged dilated convolutions and Mix Pooling.
• A hybrid CNN-Transformer model (PerceptiveNeTr) combining local and global context.
• Ablation studies, class activation maps, and qualitative/quantitative evaluations on TreeCrown, Landcover.AI, and UAVid datasets.