Research map / Architectures and efficiency / CNN design
Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CAPTN, a ConvNeXt-based texture recognition network with Texture Frequency Attention, Dual Depth Permutation, and Learnable Chebyshev Polynomials, trained with a Latent Texture Attribute loss and Stochastic Local Texture Masking.
Problem
Deep texture recognition lacks representational diversity, struggles to preserve discriminative cues across layers, and uses losses that ignore spatial coherence and statistical consistency in feature space.
Approach
Combines attention-derived latent texture attributes with channel depth permutations and a learnable Chebyshev polynomial expansion of orderless attributes, plus a three-part LTA loss (cross-entropy, orderless distribution KL alignment, and spatial patch MSE alignment) and random local patch masking.
Key results
- CAPTN with ConvNeXt-L reaches 79.6% on DTD, 93.0% on KTH, 85.0% on GTOS, and 95.1% on GTOS-M, outperforming GTN and RADAM in most settings.
- With ConvNeXt-T, LTA loss improves over cross-entropy from 78.4% to 78.9% on DTD and from 91.9% to 92.2% on KTH.
- LCP outperforms an RBF basis across all four datasets with ConvNeXt-L (e.g., GTOS-M 95.1 vs 92.7) while using fewer parameters (9.60M vs 14.22M on GTOS-M).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CAPTN in place of GTN with ConvNeXt-L | texture recognition · DTD | accuracy: 79.3 ± 0.8 -> 79.6±0.8 | Helped |
| add CAPTN in place of RADAM with ConvNeXt-B | material recognition · FMD | accuracy: 90.2±0.2 -> 89.1 ± 3.4 | Hurt |
| add CAPTN in place of GTN with ConvNeXt-L | material recognition · KTH | accuracy: 92.5 ± 4.7 -> 93.0±4.9 | Helped |
| add CAPTN in place of GTN | texture recognition · GTOS | accuracy: 84.9 ± 1.7 -> 85.1±1.7 | Helped |
| add CAPTN in place of GTN | material recognition · GTOS-M | accuracy: 94.8 -> 95.1 | Helped |
| add TFA (ablation) | CAPTN · DTD, KTH, and GTOS | accuracy: 75.4 ±1.0 -> 77.1 ±1.1 (DTD); 87.2 ±5.3 -> 90.4 ±4.8 (KTH); 81.7 ±2.0 -> 83.5 ±1.8 (GTOS) | Helped |
| add D2P to TFA (ablation) | CAPTN · DTD, KTH, and GTOS | accuracy: 77.1 ±1.1 -> 77.2 ±1.0 (DTD); 90.4 ±4.8 -> 90.5 ±4.8 (KTH); 83.5 ±1.8 -> 83.6 ±1.8 (GTOS) | Helped |
| add LCP to TFA and D2P (ablation) | CAPTN · DTD, KTH, and GTOS | accuracy: 77.2 ±1.0 -> 77.9 ±1.0 (DTD); 90.5 ±4.8 -> 90.6 ±4.6 (KTH); 83.6 ±1.8 -> 83.9 ±1.7 (GTOS) | Helped |
| use LTA Loss instead of CE Loss (ablation) | CAPTN · DTD, KTH, GTOS, and GTOS-M | accuracy: 78.4 ±0.7 -> 78.9±0.8 (DTD); 91.9 ±5.1 -> 92.2±4.4 (KTH); 85.0 ±1.8 -> 85.1±1.7 (GTOS); 93.6 -> 94.2 (GTOS-M) | Helped |
| replace first-order mean pooling with second-order covariance pooling and remove LCP (ablation) | CAPTN · DTD, KTH, GTOS, and GTOS-M | accuracy: 79.3±0.9 -> 76.4 ±1.0 (DTD); 91.3±5.7 -> 91.1 ±6.3 (KTH); 84.6±1.9 -> 83.5 ±1.7 (GTOS); 94.0 -> 93.0 (GTOS-M) | Hurt |
Surprising findings
- Second-order covariance pooling underperforms first-order mean pooling followed by LCP, which the authors attribute to noise sensitivity.
Limitations
- Reported gains over GTN are often small and within standard deviations (e.g., FMD, GTOS-M with ConvNeXt-B where GTN is higher on GTOS-M).
- Backbones are frozen, so the method does not fine-tune the feature extractor.
Experimental setup
Datasets: DTD FMD KTH-TIPS2-b GTOS GTOS-Mobile
Benchmarks: DTD FMD KTH-TIPS2-b GTOS GTOS-Mobile
Metrics: classification accuracy
Baselines: GTN RADAM CE loss RBF basis second-order covariance pooling
Builds on: GTN RADAM DeepTEN co-training
Closest papers by meaning
- Deep TEN: Texture Encoding Network (CVPR 2017 · similarity 0.86)
- Deep Structure-Revealed Network for Texture Recognition (CVPR 2020 · similarity 0.86)
- Visualizing and Understanding Deep Texture Representations (CVPR 2016 · similarity 0.85)
- MDFL: Multi-Domain Diffusion-Driven Feature Learning (AAAI 2024 · similarity 0.83)
- ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.83)
- Visual Anagrams Reveal Hidden Differences in Holistic Shape Processing Across Vision Models (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.