LCA-CNN

LCA-CNN performs fine-grained image classification by using a Learning Cascade Attention Convolutional Neural Network to distinguish highly similar classes with minimal intra-class variation.

Key Features:

• Cascade Attention Model: Employs a Deep Convolutional Neural Network (CNN) architecture with cascade attention to improve discrimination of fine-grained classes.
• Spatial Confusion Attention: Identifies ambiguous regions within input images to focus analysis on areas requiring precise differentiation.
• Constraint Loss Functions: Includes Spatial Mask Loss to penalize incorrect focus areas and Spatial And Loss to enforce consistent feature presence across class instances.
• Cross-Network Attention: Uses a two-stream architecture with distinct pre-trained parameters to leverage diverse feature representations.
• Novel Loss Functions: Implements Cross-network Similarity Loss to align features between streams and Satisfied Rank Loss to enforce prediction ranking consistent with ground truth.
• Network Fusion Attention: Integrates intermediate results using an entropy add strategy to produce final predictions.
• Weakly-supervised and Parallelized Training: Operates under a fully weakly-supervised regime and supports parallelized computation for training.
• End-to-end Training: Architecture permits end-to-end model training.

Scientific Applications:

• CUB-200-2011: Achieved a classification accuracy of 90.8% on this fine-grained bird dataset.
• FGVC-Aircraft: Attained a classification accuracy of 92.1% on this fine-grained aircraft dataset.
• Flower 102: Reached a classification accuracy of 98.5% on this fine-grained flower dataset.

Methodology:

Computational methods include a deep CNN with cascade attention and Spatial Confusion Attention, a two-stream Cross-Network Attention setup with distinct pre-trained parameters, constraint losses (Spatial Mask Loss, Spatial And Loss), Cross-network Similarity Loss, Satisfied Rank Loss, Network Fusion Attention using an entropy add strategy, and weakly-supervised, parallelized end-to-end training.