DDHTS-Net: dual-domain hierarchical texture supervision network for plant texture analysis

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Texture analysis is crucial for visual recognition but faces challenges due to subtle inter-class differences and intra-class variability in plant textures.
• Fine-grained texture variations and imaging conditions (scale, illumination, viewpoint) complicate accurate recognition.
• Existing deep learning models often use single-domain analysis, limiting their ability to capture essential cross-frequency and hierarchical cues for fine-grained discrimination.

Purpose of the Study:

• To propose a novel dual-domain hierarchical texture supervision network (DDHTS-Net) for enhanced texture analysis and classification.
• To effectively model complex texture attributes across multiple domains and hierarchical levels.
• To improve fine-grained texture discrimination in challenging scenarios, particularly for plant recognition.

Main Methods:

• The DDHTS-Net utilizes the Undecimated Wavelet Transform to create four channels with varying granularity levels.
• It incorporates an intra-channel hierarchical supervision unit where deeper feature maps guide shallower ones.
• An inter-channel granularity-level supervision unit captures distinct spatial frequency characteristics across different levels.

Main Results:

• DDHTS-Net demonstrated superior classification performance on two plant classification datasets.
• The network also achieved state-of-the-art results on two benchmark texture datasets.
• Experiments confirmed the effectiveness of dual-domain representations and hierarchical supervision for fine-grained texture discrimination.

Conclusions:

• Integrating dual-domain representations with hierarchical texture supervision significantly enhances fine-grained texture discrimination.
• The proposed DDHTS-Net offers an effective framework for complex texture attribute modeling in plant recognition and general texture classification.
• This approach provides a robust solution for challenging texture analysis tasks where subtle details are critical.