Perception-Aware Texture Similarity Prediction.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Texture similarity is crucial for material recognition but algorithmic predictions often diverge from human perception.
• Existing methods struggle with perceptually consistent fine-grained texture similarity due to misaligned representations and metrics.

Purpose of the Study:

• To develop a novel network, the Perception-Aware Texture Similarity Prediction Network (PATSP-Net), that addresses the discrepancy between algorithmic and human perception of texture similarity.
• To introduce a new approach for learning perception-aware texture representations and similarity metrics.

Main Methods:

• Introduced the Perception-Aware Texture Similarity Prediction Network (PATSP-Net).
• Developed a Bilinear Lateral Attention Transformer network (BiLAViT) incorporating a Siamese Feature Extraction Subnetwork (SFEN) and Metric Learning Subnetwork (MLN).
• Proposed a novel Ranking and Scaling Loss function (RSLoss) to measure both ranking and scaling differences.

Main Results:

• The PATSP-Net demonstrated superior or comparable performance against existing methods on three distinct fine-grained texture similarity prediction tasks.
• The proposed Bilinear Lateral Attention Transformer network (BiLAViT) and RSLoss were shown to be effective for texture similarity tasks.
• The network successfully learned perception-aware texture representations and similarity metrics.

Conclusions:

• The joint utilization of BiLAViT and RSLoss in PATSP-Net enables learning perception-aware texture representations and similarity metrics.
• PATSP-Net offers a promising solution for perceptually consistent fine-grained texture similarity prediction.
• This work advances texture analysis by bridging the gap between computational and human perception.