Graph Jigsaw Learning for Cartoon Face Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Cartoon face recognition is difficult due to smooth regions and emphasized edges, requiring precise perception of sparse shape patterns.
• Convolutional Neural Networks (CNNs) struggle to learn shape-oriented representations for cartoon faces effectively.
• Limited texture information in cartoon faces necessitates a focus on shape pattern recognition.

Purpose of the Study:

• To propose GraphJigsaw, a novel method for enhancing cartoon face recognition by learning shape patterns.
• To leverage graph convolutional networks (GCN) for self-supervised recovery of spatial feature map layouts.
• To integrate shape-pattern learning progressively within classification networks without increasing inference complexity.

Main Methods:

• GraphJigsaw constructs jigsaw puzzles by spatially shuffling intermediate convolutional feature maps at various network stages.
• A graph convolutional network (GCN) is employed to progressively solve these puzzles in a self-supervised manner.
• The method avoids training with deconstructed images, preventing noisy pattern introduction and focusing on shape recovery.

Main Results:

• GraphJigsaw effectively learns and propagates shape patterns throughout the classification model.
• The approach requires no additional manual annotations and imposes no extra computational burden during inference.
• Experimental results demonstrate significant performance improvements over existing face recognition and jigsaw-based methods on cartoon face datasets.

Conclusions:

• GraphJigsaw offers a feasible and effective solution for the challenging problem of cartoon face recognition.
• The self-supervised, shape-oriented approach using GCNs provides a robust representation learning strategy.
• The method's seamless integration and efficiency make it a valuable contribution to the field of computer vision.