Improvements to neural networks have driven advances in many diverse fields. Traditionally, these enhancements have focused on refining internal structures or training procedures. By viewing neural networks as approximations of the human neural model, an alternative strategy emerges: designing architectures that closely mirror human biology. This paper introduces new architectures that learn human-based low-level visual features, grounded in neuroscience concepts, for segmentation tasks. Initially, we train U-Net-like architectures to develop a human-like perception space representation. This method leverages neuroscience insights, incorporating luminance and chrominance from the YUV space, and line orientation detected via a probabilistic Hough method. Subsequently, a transfer learning strategy is employed, using these pre-trained architectures as foundations for novel U-Net-like segmentation models. Evaluations on the CamVid dataset demonstrate that these models achieve performance comparable to well-established methods while utilizing simpler architectures, suggesting potential performance improvements through neuroscience-inspired designs.
Human-Based Low-Level Visual Processing Neural Network for Image Segmentation
Fagioli A.
;Foresti G. L.;
2025-01-01
Abstract
Improvements to neural networks have driven advances in many diverse fields. Traditionally, these enhancements have focused on refining internal structures or training procedures. By viewing neural networks as approximations of the human neural model, an alternative strategy emerges: designing architectures that closely mirror human biology. This paper introduces new architectures that learn human-based low-level visual features, grounded in neuroscience concepts, for segmentation tasks. Initially, we train U-Net-like architectures to develop a human-like perception space representation. This method leverages neuroscience insights, incorporating luminance and chrominance from the YUV space, and line orientation detected via a probabilistic Hough method. Subsequently, a transfer learning strategy is employed, using these pre-trained architectures as foundations for novel U-Net-like segmentation models. Evaluations on the CamVid dataset demonstrate that these models achieve performance comparable to well-established methods while utilizing simpler architectures, suggesting potential performance improvements through neuroscience-inspired designs.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
