Edge-Aware Loss Function for Small-Object Pixel Classification

The patent applies local quality by differentiating the loss function treatment for different regions of the image. Boundary pixels and pixels from small objects are identified and assigned higher weights in the loss function, while other pixels use standard weighting. This allows the training process to focus computational attention on critical regions that require higher classification accuracy, resolving the contradiction between training simplicity and classification accuracy for small objects and boundaries.

The patent changes the parameter weighting in the loss function dynamically based on pixel characteristics. Instead of uniform weighting, the loss function incorporates adaptive weights that increase for boundary pixels and pixels belonging to small objects. This parameter change maintains computational efficiency while significantly improving segmentation reliability for critical regions by making the loss landscape more sensitive to errors in important areas.

The patent applies preliminary action by pre-identifying boundary pixels and small object pixels before the main training process. These critical pixels are marked and assigned special weights in advance, allowing the standard CNN training algorithm to proceed at high speed while automatically focusing on the right regions. This preliminary classification enables the model to achieve high boundary detection precision without sacrificing processing speed, as the weighted loss function guides learning toward critical regions during normal training operations.