Easy2Hard

Easy2Hard enhances structure-preserving image smoothing using deep learning to remove unimportant details while retaining essential edges and structures for computer vision and graphics applications.

Key Features:

• Synthetic Dataset Generation: Generates a synthetic dataset that semantically separates meaningful structures from unimportant details, comprising "easy" samples with ground-truth labels from candidate generation and screening and "hard" samples synthesized via structure-preserving smoothing.
• Joint Edge Detection and Smoothing Network (JESS-Net): Implements JESS-Net to perform joint edge detection and structure-preserving image smoothing within a single neural network.
• Total Variation Loss: Incorporates total variation loss as prior knowledge to reduce the gap between synthetic training data and real-world images, improving generalization across datasets.

Scientific Applications:

• Computer vision and graphics: Structure-preserving image enhancement and preprocessing for computer vision and graphics tasks.
• Medical imaging: Noise reduction while preserving critical anatomical edges in medical images.
• Satellite imagery analysis: Smoothing remote sensing images while maintaining structural features relevant to analysis.

Methodology:

Creates a synthetic dataset by producing "easy" samples with ground-truth via candidate generation and screening and "hard" samples via structure-preserving smoothing, then trains JESS-Net on this dataset using total variation loss and evaluates performance on various datasets and real images.