Edge-Reinforced Image Datasets for CNN Hazard Detection

The patent applies local quality by assigning different weights to different regions of the image, specifically emphasizing edge regions. The loss function uses weight maps that give higher weights to edge pixels and lower weights to non-edge pixels, allowing the model to focus computational attention on critical edge detection while maintaining overall segmentation performance.

The patent implements preliminary action by pre-processing the input image to generate edge maps and weight maps before feeding them into the segmentation model. This pre-computed edge information is incorporated into the loss function, guiding the model to prioritize edge region accuracy from the beginning of the training process rather than relying solely on the model to discover edges during training.

The patent introduces intermediary components including pre-trained edge detection models (such as Canny or Sobel operators) and weight map generators that serve as mediators between the input image and the segmentation model. These intermediaries process the image to extract edge features and generate weight distributions, which then guide the segmentation model without requiring fundamental changes to the model architecture itself.

The patent segments the training process into distinct components: edge detection preprocessing, weight map generation, weighted loss calculation, and model training. This segmentation allows each component to be optimized independently and enables flexible adjustment of edge reinforcement strength without retraining the entire system from scratch.