Hadamard Layer for Improve Semantic Segmentation

Semantic segmentation is crucial in image analysis but remains challenging due to performance and robustness issues. This paper introduces the Hadamard Layer, an enhancement tested in the Pix2Pix encoder models that improve segmentation tasks using a novel class encoding scheme. This scheme increases the Hamming distance between class labels using Hadamard codes instead of classic one-hot encoding, thus enhancing model robustness against adversarial attacks and improving segmentation precision. The Hadamard Layer was integrated into various UNet variations encoder models and tested across multiple datasets, including medical images, urban scenes, and facil datasets. Our evaluations show substantial improvements in pixel accuracy and Intersection over Union (IoU) metrics without increasing the model’s parameter count. Specifically, models with the Hadamard Layer showed up to a 3% increase in pixel accuracy and an 8% increase in Intersection over Union (IoU) on the LiTS dataset. On the CelebA dataset, models saw an average improvement of +18% in pixel accuracy and +9% in IoU. Pixel accuracy increased by 7% for the CITYSCAPES dataset and IoU by 4% across most models. The results affirm the Hadamard Layer’s potential to enhance semantic segmentation performance significantly, setting a foundation for future research in more complex applications.