DeriveNet for (Very) Low Resolution Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Image Processing

Background:

• Very low resolution (VLR/LR) images pose challenges for automated identification due to limited information content.
• Traditional feature extraction and classification methods are often ineffective for VLR/LR data.

Purpose of the Study:

• To propose a novel DeriveNet model for effective VLR/LR classification.
• To enhance classification by learning effective class boundaries using class-specific domain knowledge.

Main Methods:

• DeriveNet is jointly trained using Derived-Margin softmax loss and Reconstruction-Center (ReCent) loss.
• Derived-Margin softmax loss models inter-class variations for VLR/LR classification.
• ReCent loss learns a high-resolution reconstruction space to approximate class variations, informing the Derived-Margin softmax loss.

Main Results:

• The DeriveNet model incorporates Multi-resolution Pyramid based data augmentation for training across varying resolutions.
• Experiments were conducted on VLR/LR face recognition and digit classification tasks, including drone-shot videos.
• DeriveNet achieved state-of-the-art performance on multiple datasets.

Conclusions:

• The DeriveNet model demonstrates significant utility and effectiveness for various VLR/LR classification applications.
• The proposed losses and data augmentation strategy contribute to improved VLR/LR image analysis.
• This research advances automated identification capabilities for degraded image quality scenarios.