pith. sign in

arxiv: 2605.05012 · v1 · submitted 2026-05-06 · 💻 cs.CV

Chaotic Contrastive Learning for Robust Texture Classification

Joao B Florindo This is my paper

Pith reviewed 2026-05-08 18:13 UTC · model grok-4.3

classification 💻 cs.CV
keywords texture classificationchaotic mapscontrastive learningself-supervised learningdata augmentationcomputer visionrobust features
0
0 comments X

The pith

Pixel-wise chaotic maps from deterministic dynamics serve as non-linear augmentations in contrastive pre-training to learn topologically robust texture features.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes a self-supervised framework that applies Logistic, Tent, and Sine chaotic maps at the pixel level as data augmentations during contrastive pre-training. These perturbations, drawn from ergodic theory, are intended to simulate complex environmental noise and reflectance changes so the network learns features invariant to scale, illumination, and structural variations common in textures. An attention-based ensemble then combines low-frequency structural details from the small chaos-pretrained encoder with high-level semantic features from a larger supervised backbone. Experiments on the FMD, UMD, KTH-TIPS2-b, DTD, GTOS, and 1200Tex datasets show the method surpasses prior state-of-the-art results.

Core claim

The central claim is that pixel-wise chaotic maps (Logistic, Tent, and Sine) function as non-linear data augmentations within a contrastive self-supervised pre-training stage, forcing the model to extract topologically robust features that mimic real-world noise and reflectance variations, and that fusing these with semantic representations via attention yields higher accuracy than existing approaches across six texture benchmarks.

What carries the argument

Pixel-wise chaotic maps (Logistic, Tent, Sine) used as non-linear augmentations in contrastive SSL, grounded in ergodic theory to enforce topological robustness in learned features.

If this is right

  • The approach achieves higher accuracy than prior methods on the FMD, UMD, KTH-TIPS2-b, DTD, GTOS, and 1200Tex texture datasets.
  • The network becomes less reliant on color and shape cues and more invariant to scale and illumination shifts.
  • The attention-based ensemble successfully integrates low-frequency structural features from the chaos-pretrained encoder with high-level semantics.
  • The method works without requiring large amounts of labeled texture data during the pre-training stage.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar pixel-level chaotic perturbations could be tested in other image domains that suffer from uncontrolled noise, such as medical or satellite imagery.
  • The framework might reduce the volume of labeled examples needed for training texture classifiers in industrial inspection tasks.
  • Substituting different chaotic systems or varying the map parameters offers a direct way to measure how ergodic properties influence feature robustness.

Load-bearing premise

Chaotic perturbations generated by the Logistic, Tent, and Sine maps effectively replicate complex environmental noise and reflectance variations in a manner that produces topologically robust features.

What would settle it

Replacing the chaotic maps with conventional random augmentations inside the identical contrastive pre-training pipeline and observing equal or higher accuracy on the same six benchmarks would falsify the claimed benefit of the chaotic mechanism.

Figures

Figures reproduced from arXiv: 2605.05012 by Joao B Florindo.

Figure 1
Figure 1. Figure 1: Proposed chaotic contrastive learning methodology. Stage 1 learns structural texture view at source ↗
Figure 2
Figure 2. Figure 2: Confusion Matrices for the FMD dataset using different numbers of pretraining epochs (15 view at source ↗
Figure 3
Figure 3. Figure 3: Confusion Matrices for the large-scale benchmarks using the Sine Map configuration. view at source ↗
Figure 4
Figure 4. Figure 4: Confusion Matrix for the 1200Tex application task. view at source ↗
read the original abstract

Texture classification is a pivotal task in computer vision, presenting unique challenges due to high inter-class similarity and the sensitivity of structural patterns to scale and illumination changes. While Convolutional Neural Networks (CNNs) and recent Vision Transformers have set performance benchmarks, they often require extensive labeled datasets or struggle to generalize across domains due to an over-reliance on color and shape features. This paper introduces a novel framework that synergizes Self-Supervised Learning (SSL) with deterministic chaotic dynamics. We propose a chaotic contrastive pre-training strategy, where pixel-wise chaotic maps, specifically Logistic, Tent, and Sine maps, act as non-linear data augmentation techniques. These chaotic perturbations, grounded in ergodic theory, force the network to learn topologically robust features by mimicking complex environmental noise and reflectance variations. Furthermore, we introduce an attention-based feature ensemble that fuses high-level semantic representations from a supervised large backbone with low-frequency structural features from a chaos-pretrained tiny encoder. Experimental results on six texture benchmarks (FMD, UMD, KTH-TIPS2-b, DTD, GTOS, and 1200Tex) demonstrate the superiority of the proposed method, outperforming state-of-the-art approaches and achieving promising accuracies on all the analyzed datasets.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper introduces a chaotic contrastive pre-training strategy using pixel-wise Logistic, Tent, and Sine maps as non-linear data augmentations grounded in ergodic theory to learn topologically robust features for texture classification. It combines this with an attention-based feature ensemble fusing a supervised large backbone and a chaos-pretrained tiny encoder, claiming to outperform state-of-the-art on six texture benchmarks: FMD, UMD, KTH-TIPS2-b, DTD, GTOS, and 1200Tex.

Significance. If the chaotic augmentations demonstrably improve topological robustness beyond standard contrastive learning and the ensemble, the approach could advance robust texture classification under noise and illumination changes. The multi-dataset empirical evaluation is a positive aspect, but without isolating controls the contribution of the ergodic component remains unclear.

major comments (3)
  1. Abstract: The superiority claim on six datasets is asserted without reference to experimental protocol, baselines, statistical tests, ablation studies, or error bars, preventing assessment of whether the data support the central claims.
  2. Method (chaotic pre-training description): The claim that pixel-wise chaotic maps force topologically robust features via ergodic theory lacks any derivation linking ergodicity to feature topology and provides no comparison to non-chaotic noise with matched statistics.
  3. Experiments: No ablation is reported that removes the chaotic maps while retaining the contrastive loss and attention ensemble, which is required to show that the chaotic pre-training is load-bearing for the headline performance gains rather than the backbone or fusion alone.
minor comments (1)
  1. Abstract: The phrase 'promising accuracies' is imprecise; quantitative results or relative improvements should be stated explicitly even in the abstract.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment point by point below, indicating where we agree revisions are needed and how we will strengthen the manuscript.

read point-by-point responses

  1. Referee: Abstract: The superiority claim on six datasets is asserted without reference to experimental protocol, baselines, statistical tests, ablation studies, or error bars, preventing assessment of whether the data support the central claims.

    Authors: We agree that the abstract is too concise and does not reference the supporting experimental details. The full manuscript describes the evaluation protocol (standard train/test splits and metrics for each benchmark), lists all baselines, reports ablation studies on map types and ensemble components, and includes mean accuracies with standard deviations across runs in the result tables. We will revise the abstract to briefly note the six-benchmark evaluation, comparisons to SOTA methods, and the use of statistical reporting, while respecting length constraints. revision: yes

  2. Referee: Method (chaotic pre-training description): The claim that pixel-wise chaotic maps force topologically robust features via ergodic theory lacks any derivation linking ergodicity to feature topology and provides no comparison to non-chaotic noise with matched statistics.

    Authors: The current text invokes ergodicity to justify dense coverage of the input space by the maps, which we argue promotes invariance to local perturbations akin to illumination and scale changes in textures. We acknowledge that an explicit derivation connecting ergodic properties to topological feature robustness is absent. We will add a concise paragraph in Section 3 deriving the link from the dense-orbit property of ergodic maps to learned invariance, and we will include a new ablation comparing the chaotic maps against additive Gaussian noise with matched first- and second-order statistics to isolate the chaotic component. revision: partial

  3. Referee: Experiments: No ablation is reported that removes the chaotic maps while retaining the contrastive loss and attention ensemble, which is required to show that the chaotic pre-training is load-bearing for the headline performance gains rather than the backbone or fusion alone.

    Authors: We agree this control is necessary. Existing ablations vary the chaotic map family and the fusion module but do not replace the chaotic augmentations with standard contrastive augmentations while keeping the loss and ensemble fixed. We will add this experiment in the revised version, training an otherwise identical model with conventional augmentations (random crop, flip, color jitter) in the pre-training stage and reporting the resulting accuracy drop on the six benchmarks. revision: yes

Circularity Check

0 steps flagged

No derivation chain; purely empirical claims with external benchmarks

full rationale

The paper presents an empirical framework for texture classification via chaotic contrastive pre-training and attention-based ensemble. No equations, derivations, or first-principles results are described in the provided abstract or reader summary. The appeal to ergodic theory is an asserted grounding for why chaotic maps (Logistic/Tent/Sine) produce topologically robust features, but this is not derived or shown to reduce to any internal input by construction. Performance superiority is claimed via direct comparison to SOTA on six external benchmarks (FMD, UMD, etc.), which are falsifiable outside the paper and not forced by any fitted parameter renamed as prediction. No self-citation load-bearing steps, uniqueness theorems, or ansatz smuggling are evident. The work is self-contained as an experimental study; any circularity would require explicit equations or self-referential reductions that are absent here.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no technical equations, parameters, or proofs, so no free parameters, axioms, or invented entities can be identified.

pith-pipeline@v0.9.0 · 5506 in / 1121 out tokens · 61538 ms · 2026-05-08T18:13:46.439729+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

46 extracted references · 5 canonical work pages

  1. [1]

    Narayan, P

    V. Narayan, P. K. Mall, S. Awasthi, S. Srivastava, A. Gupta, Fuzzynet: medical image classification based on glcm texture feature, in: 2023 international confer- ence on artificial intelligence and smart communication (AISC), IEEE, 2023, pp. 769–773. 16

    2023

  2. [2]

    J. Yi, J. Mao, H. Zhang, K. Zeng, Z. Tao, H. Zhong, S. Wang, Y. Wang, Pstl-net: a patchwise self-texture-learning network for transmission line inspection, IEEE Transactions on Instrumentation and Measurement 73 (2023) 1–14

    2023

  3. [3]

    Akiva, M

    P. Akiva, M. Purri, M. Leotta, Self-supervised material and texture representation learning for remote sensing tasks, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 8203–8215

    2022

  4. [4]

    Z. Liu, R. Xue, Medical image encryption using biometric image texture fusion, Journal of Medical Systems 47 (1) (2023) 112

    2023

  5. [5]

    Aggarwal, M

    A. Aggarwal, M. Kumar, Image surface texture analysis and classification using deep learning, Multimedia Tools and Applications 80 (1) (2021) 1289–1309

    2021

  6. [6]

    Ojala, M

    T. Ojala, M. Pietikainen, T. Maenpaa, Multiresolution gray-scale and rotation invariant texture classification with local binary patterns, IEEE Transactions on pattern analysis and machine intelligence 24 (7) (2002) 971–987

    2002

  7. [7]

    Dinstein, K

    I. Dinstein, K. Shanmugam, R. Haralick, Textural features for image classification, IEEE Transactions on Systems, Man, and Cybernetics 3 (6) (1973) 610–621

    1973

  8. [8]

    Scabini, A

    L. Scabini, A. Sacilotti, K. M. Zielinski, L. C. Ribas, B. De Baets, O. M. Bruno, A comparative survey of vision transformers for feature extraction in texture analysis, Journal of Imaging 11 (9) (2025) 304

    2025

  9. [9]

    S. Bell, P. Upchurch, N. Snavely, K. Bala, Material recognition in the wild with the materials in context database, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 3479–3487

    2015

  10. [10]

    J. Gui, T. Chen, J. Zhang, Q. Cao, Z. Sun, H. Luo, D. Tao, A survey on self-supervised learning: Algorithms, applications, and future trends, IEEE Transactions on Pattern Analysis and Machine Intelligence 46 (12) (2024) 9052– 9071

    2024

  11. [11]

    T. Chen, S. Kornblith, M. Norouzi, G. Hinton, A simple framework for contrastive learning of visual representations, in: Proceedings of the 37th International Conference on Machine Learning (ICML), 2020, pp. 1597–1607

    2020

  12. [12]

    N. Chen, Z. Xu, Z. Liu, Y. Chen, Y. Miao, Q. Li, Y. Hou, L. Wang, Data augmentation and intelligent recognition in pavement texture using a deep learning, IEEE Transactions on Intelligent Transportation Systems 23 (12) (2022) 25427–25436. 17

    2022

  13. [13]

    R. M. May, Simple mathematical models with very complicated dynamics, Nature 261 (5560) (1976) 459–467

    1976

  14. [14]

    M. T. Elkandoz, W. Alexan, Image encryption based on a combination of multiple chaotic maps, Multimedia Tools and Applications 81 (18) (2022) 25497–25518

    2022

  15. [15]

    S. Woo, S. Debnath, R. Hu, X. Chen, Z. Liu, I. S. Kweon, S. Xie, Convnext v2: Co-designing and scaling convnets with masked autoencoders, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 16133–16142

    2023

  16. [16]

    J. Hu, L. Shen, G. Sun, Squeeze-and-excitation networks, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 7132–7141

    2018

  17. [17]

    G. V. de Lima, P. T. Saito, F. M. Lopes, P. H. Bugatti, Classification of texture based on bag-of-visual-words through complex networks, Expert Systems with Applications 133 (2019) 215–224

    2019

  18. [18]

    Zhang, J

    H. Zhang, J. Xue, K. Dana, Deep ten: Texture encoding network, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 708–717

    2017

  19. [19]

    Y. Xu, F. Li, Z. Chen, J. Liang, Y. Quan, Encoding spatial distribution of convolutional features for texture representation, Advances in Neural Information Processing Systems 34 (2021) 22732–22744

    2021

  20. [20]

    Z. Chen, F. Li, Y. Quan, Y. Xu, H. Ji, Deep texture recognition via exploiting cross-layer statistical self-similarity, in: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, 2021, pp. 5231–5240

    2021

  21. [21]

    M. Tian, L. Tang, J. Xu, Y. Zhang, Y. Yang, L. Zeng, E. Chen, Y. Xie, Hybrid cnn-transformer framework with dynamic feature fusion for enhanced passport background texture classification: M. tian et al., The Visual Computer 42 (1) (2026) 4

    2026

  22. [22]

    K. He, H. Fan, Y. Wu, S. Xie, R. Girshick, Momentum contrast for unsupervised visual representation learning, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 9729–9738. 18

    2020

  23. [23]

    K. He, X. Chen, S. Xie, Y. Li, P. Dollár, R. Girshick, Masked autoencoders are scalable vision learners, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 16000–16009

    2022

  24. [24]

    X. Liu, F. Zhang, Z. Hou, L. Mian, Z. Wang, J. Zhang, J. Tang, Self-supervised learning: Generative or contrastive, IEEE transactions on knowledge and data engineering 35 (1) (2021) 857–876

    2021

  25. [25]

    R. A. Bafghi, N. Harilal, C. Monteleoni, M. Raissi, MixDiff: Mixing natural and synthetic images for robust self-supervised representations, in: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), IEEE, 2025, pp. 7500–7511

    2025

  26. [26]

    J. Lu, X. Xia, X. Zhang, R. Zhao, Y. Zhang, Multiple-image encryp- tion algorithm based on a new 3d hyperchaotic map and whac-a-mole scrambling model, Expert Systems with Applications 290 (2025) 128393. doi:https://doi.org/10.1016/j.eswa.2025.128393. URL https://www.sciencedirect.com/science/article/pii/ S0957417425020123

    work page doi:10.1016/j.eswa.2025.128393 2025

  27. [27]

    B. Jia, Z. Guo, T. Huang, F. Guo, H. Wu, A generalized lorenz system-based initialization method for deep neural networks, Applied Soft Computing 167 (2024) 112316

    2024

  28. [28]

    Sharan, R

    L. Sharan, R. Rosenholtz, E. H. Adelson, Accuracy and speed of material categorization in real-world images, Journal of Vision 14 (9) (2014) 12–12. doi:10.1167/14.9.12

    work page doi:10.1167/14.9.12 2014

  29. [29]

    Y. Xu, H. Ji, C. Fermüller, Viewpoint invariant texture description using fractal analysis, International Journal of Computer Vision 83 (1) (2009) 85–100.doi: 10.1007/s11263-009-0220-6

    work page doi:10.1007/s11263-009-0220-6 2009

  30. [30]

    Caputo, E

    B. Caputo, E. Hayman, P. Mallikarjuna, Class-specific material categorisation, in: Proceedings of the Tenth IEEE International Conference on Computer Vision (ICCV), Vol. 2, IEEE, 2005, pp. 1597–1604

    2005

  31. [31]

    Casanova, J

    D. Casanova, J. J. de Mesquita Sá Junior, O. M. Bruno, Plant leaf identification using gabor wavelets, International Journal of Imaging Systems and Technology 19 (3) (2009) 236–243.doi:10.1002/ima.20201. 19

    work page doi:10.1002/ima.20201 2009

  32. [32]

    Cimpoi, S

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, A. Vedaldi, Describing textures in the wild, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014, pp. 3606–3613

    2014

  33. [33]

    J. Xue, H. Zhang, K. Dana, K. Nishino, Differential angular imaging for material recognition, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 764–773

    2017

  34. [34]

    Mircea, M

    C. Mircea, M. Subhransu, A. Vedaldi, Deep filter banks for texture recognition, description, and segmentation, International Journal of Computer Vision (2016) 65–94

    2016

  35. [35]

    Y. Song, F. Zhang, Q. Li, H. Huang, L. J. O’Donnell, W. Cai, Locally-transferred fisher vectors for texture classification, in: Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 4912–4920

    2017

  36. [36]

    Jbene, A

    M. Jbene, A. D. El Maliani, M. El Hassouni, Fusion of convolutional neural network and statistical features for texture classification, in: 2019 International Conference on Wireless Networks and Mobile Communications (WINCOM), IEEE, 2019, pp. 1–4

    2019

  37. [37]

    W. Zhai, Y. Cao, Z.-J. Zha, H. Xie, F. Wu, Deep structure-revealed network for texture recognition, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 11010–11019

    2020

  38. [38]

    J. B. Florindo, Y.-S. Lee, K. Jun, G. Jeon, M. K. Albertini, Visgraphnet: A complex network interpretation of convolutional neural features, Information Sciences 543 (2021) 296–308

    2021

  39. [39]

    Florindo, K

    J. Florindo, K. Metze, Using non-additive entropy to enhance convolutional neural features for texture recognition, Entropy 23 (10) (2021) 1259

    2021

  40. [40]

    S. Mao, D. Rajan, L. T. Chia, Deep residual pooling network for texture recog- nition, Pattern Recognition 112 (2021) 107817

    2021

  41. [41]

    Z. Yang, S. Lai, X. Hong, Y. Shi, Y. Cheng, C. Qing, Dfaen: Double-order knowledge fusion and attentional encoding network for texture recognition, Expert Systems with Applications 209 (2022) 118223

    2022

  42. [42]

    Scabini, K

    L. Scabini, K. M. Zielinski, L. C. Ribas, W. N. Gonçalves, B. De Baets, O. M. Bruno, Radam: Texture recognition through randomized aggregated encoding of deep activation maps, Pattern Recognition 143 (2023) 109802. 20

    2023

  43. [43]

    Mamidibathula, S

    B. Mamidibathula, S. Amirneni, S. S. Sistla, N. Patnam, Texture classification using capsule networks, in: Iberian Conference on Pattern Recognition and Image Analysis, Springer, 2019, pp. 589–599

    2019

  44. [44]

    L. O. Lyra, A. E. Fabris, J. B. Florindo, A multilevel pooling scheme in convolu- tional neural networks for texture image recognition, Applied Soft Computing 152 (2024) 111282.doi:10.1016/j.asoc.2023.111282

    work page doi:10.1016/j.asoc.2023.111282 2024

  45. [45]

    P. M. Silva, J. B. Florindo, Fractal measures of image local features: An appli- cation to texture recognition, Multimedia Tools and Applications 80 (9) (2021) 14213–14229

    2021

  46. [46]

    J. B. Florindo, E. E. Laureano, Boff: A bag of fuzzy deep features for texture recognition, Expert Systems with Applications 219 (2023) 119627. 21

    2023