Pith. sign in

REVIEW 3 major objections 5 minor 61 references

Feature-Enhanced TResNet for Fine-Grained Food Image Classification

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A convolutional attention upgrade to TResNet-XL claims 81.37% Top-1 accuracy on ChineseFoodNet.

desk verdict A small incremental CNN tweak with an internally consistent ablation, but the headline claim of 'significantly outperform' rests on an unfair and undocumented baseline comparison. read the letter →

arxiv 2507.12828 v2 pith:T7XQ7LCE submitted 2025-07-17 cs.CV cs.AI

classification cs.CVcs.AI
keywords fine-grainedfoodclassificationFE-TResNetStyle-basedRecalibrationModuleDeepChannel-wiseAttentionTResChineseCNFOOD-241precisionnutrition
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes FE-TResNet, a convolutional network for fine-grained food recognition built on the TResNet-XL backbone. Its central claim is that inserting two modules — a Style-based Recalibration Module (StyleRM) and a Deep Channel-wise Attention module (DCA) — raises Top-1 accuracy to 81.37% on ChineseFoodNet and 80.29% on CNFOOD-241, above every baseline the authors compare against. The reported ablation attributes most of the Top-1 gain to StyleRM, with DCA adding its clearest contribution to Top-5 accuracy when combined. If the gains hold under a shared training protocol, the model gives precision-nutrition systems a concrete CNN route to telling visually similar dishes apart.

What carries the argument

The load-bearing components are StyleRM and DCA, inserted into TResNet-XL's residual blocks. StyleRM pools each channel's mean and standard deviation over spatial positions, encodes these style statistics with a convolution, batch-normalization, and activation stack, and reweights the feature map so texture and style cues are emphasized before the Squeeze-and-Excitation (SE) block. DCA, implemented as depthwise-separable criss-cross attention, uses depthwise-separable 1x1 convolutions to form query, key, and value maps, computes affinities along each pixel's row and column, runs two parameter-shared passes to approximate global context, and adds the aggregated result back to the features. The backbone TResNet-XL contributes high-throughput residual blocks and pretrained weights; the two modules are the claimed source of the accuracy increase.

What would settle it

Retrain the Table 2 baselines on ChineseFoodNet and CNFOOD-241 using the FE-TResNet protocol — ImageNet-21K pretraining, Adam at 1e-4, cosine decay, AutoAugment, 100 epochs, batch size 48, 224-pixel input — and compare Top-1 accuracy; if TResNet-XL or EfficientNetV2-S then matches or beats 81.37% and 80.29%, the central claim is refuted. Rerunning the Table 3 ablation across multiple seeds would also check whether the StyleRM gain is stable.

Watch

Extended reading notes

Core claim

The discovery the authors report is that feature enhancement, rather than a new backbone, is what moves fine-grained food accuracy. Grafting StyleRM and DCA onto TResNet-XL yields 81.37% Top-1 and 97.86% Top-5 on ChineseFoodNet, and 80.29% Top-1 and 97.97% Top-5 on CNFOOD-241, compared with 80.85% and 79.85% Top-1 for the unmodified TResNet-XL on the two datasets. The Table 3 ablation shows that StyleRM alone supplies most of the Top-1 improvement (81.22% and 80.15%), while DCA alone moves Top-1 little (80.91% and 79.92%); together they produce the best Top-5 scores. The paper interprets this as better handling of intra-class variability and inter-class similarity in Chinese food images.

Load-bearing premise

The comparison assumes every baseline model was trained with the same effective recipe as FE-TResNet, but the paper gives full training details only for its own model and the baselines run at different input resolutions, so the accuracy gap could come from training effort rather than the new modules.

Editorial extensions

If this is right

  • If correct, FE-TResNet improves Top-1 accuracy by about 0.5 points over unmodified TResNet-XL on ChineseFoodNet and about 0.4 points on CNFOOD-241, with larger Top-5 gains of roughly one and two points, respectively.
  • The modules add about 6.6 million parameters and 0.65 G FLOPs over TResNet-XL, so the reported gain does not require a new backbone or a fundamentally larger model.
  • StyleRM is the primary driver of Top-1 accuracy, suggesting texture and style reweighting is the mechanism that matters most for these Chinese food datasets.
  • For dietary monitoring, the Top-5 gains mean the correct dish is more often in the model's shortlist, which is the practical setting where a user confirms a recommendation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the StyleRM ablation suggests texture, not just shape, carries much of the discriminative signal in Chinese food; a natural test is to apply FE-TResNet to other texture-dominated fine-grained domains and measure whether the style-module gain transfers.
  • Because the paper does not document the baselines' training recipe, a controlled re-run under the FE-TResNet protocol is needed to know whether the reported margins are architectural or reflect undertrained baselines; this is a fair head-to-head the authors do not provide.
  • Since CNFOOD-241 was built from ChineseFoodNet, the two evaluations are not fully independent; testing on a different cuisine or a general fine-grained benchmark would define how far the claim extends.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes FE-TResNet, a TResNet-based architecture for fine-grained Chinese food image classification. It augments TResNet-XL with two modules: a Style-based Recalibration Module (StyleRM), which extracts channel-wise style statistics and recalibrates features with a small convolutional branch, and Deep Channel-wise Attention (DCA), a criss-cross attention variant using depthwise separable convolutions for multi-scale feature fusion. Experiments on ChineseFoodNet and CNFOOD-241 report Top-1 accuracies of 81.37% and 80.29%, with an ablation showing that StyleRM contributes most of the gain and DCA a smaller additional gain. The paper claims these results significantly outperform existing techniques and positions the model for dietary assessment and precision nutrition applications.

Significance. If the reported results are reproducible and the comparisons fair, the contribution is a modest but plausible architecture-level improvement for a practical fine-grained classification task. The ablation is internally coherent and usefully reports parameter counts, FLOPs, and memory usage, and it shows a parameter-free gain from StyleRM. The authors should be credited for testing on two public large-scale datasets and for providing a modular ablation. However, the significance of the work currently rests on the claim of significant superiority over existing methods; that claim is not yet established because the comparison omits the food-specific state-of-the-art models cited in the introduction and because the baseline training protocol is not documented. The reported top-1 margins over the closest generic baseline are small, and the ablation is presented without variance information.

major comments (3)
  1. [§1 and §4.3 (Table 2)] The central claim that FE-TResNet's accuracies "significantly outperform existing techniques" is not supported by the comparison presented. Table 2 includes only generic backbone models and omits the food-specific methods cited in the introduction — FoodCSwin [4], AlsmViT [31], FoodSwin [32], and FGFoodNet [34]. If any of these already reports a higher Top-1 accuracy on ChineseFoodNet or CNFOOD-241, the claimed superiority fails; the paper must either include these methods in the comparison with documented training protocols or qualify the claim to "among the compared generic backbones." The narrow margin over EfficientNetV2-S (0.26% on ChineseFoodNet) makes this omission consequential.
  2. [§4.1, §4.3, Table 1] The fairness of the comparison is not established. Section 4.3 states "consistent configurations were maintained," but Table 1 documents a training recipe only for FE-TResNet; the baselines' optimizers, learning-rate schedules, augmentation, and pretraining are unspecified. Differences in Table 2 could therefore reflect training effort rather than architecture. In addition, the caption "50 rounds of best results" suggests selection of the best test performance over runs; the paper should specify whether the validation or test set was used for this selection and report standard deviations, otherwise the reported margins risk being inflated by test-set overfitting.
  3. [§4.4, Table 3] The ablation supports the descriptive claim that StyleRM and DCA each contribute, but the magnitudes are small: StyleRM adds +0.37 and +0.30 Top-1 points, DCA adds +0.06 and +0.07 points, with no variance or number of seeds reported. A single-run difference of 0.06 percentage points on a test set of thousands of images is within plausible binomial noise, so the DCA contribution and even the incremental gain of StyleRM should be treated as preliminary. The paper should report multiple runs with mean and standard deviation (or confidence intervals) for at least Table 2 and Table 3 before claiming significance.
minor comments (5)
  1. [§3.3 and §4.2] Equations (1)–(6) contain inconsistent notation (e.g., b c appears both as a subscript and as an index, and the equality symbols around pooled statistics are garbled); all symbols should be defined before use. Equations (10)–(11) use P1, P5, and PA, but these are defined only in the paragraph below the equations; the definitions should precede the formulas.
  2. [Table 2] There are typos in the table: "MoblienetV2," "MoblienetV3-s," and "MoblienetV3-l" should read "MobileNetV2" etc., and "Resnet" should be "ResNet" throughout.
  3. [§5] The conclusion states the model was compared with "other self-supervised models," but FE-TResNet and the listed baselines are all trained with full supervision; this characterization should be corrected.
  4. [References] References [38], [39], and [42] are the same paper (Dichao Liu et al., "Learn from each other to classify better") cited three times; this should be consolidated. Also, the text at Section 2.2 cites [45,46,47] for style transfer but the surrounding discussion of CNNs and texture bias would be better supported by [48,49], which are cited only later.
  5. [Table 4] The F1, Precision, and Recall values are reported without specifying whether they are macro-averaged, micro-averaged, or class-weighted; since ChineseFoodNet is class-imbalanced, this choice materially affects the numbers and should be stated.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the reported accuracies are measured on held-out splits, not derived by construction; the main weakness is missing food-specific baselines, which is a comparison-fairness issue rather than circular reasoning.

full rationale

FE-TResNet's reported Top-1 and Top-5 accuracies are the outcome of supervised training and evaluation on public test splits of ChineseFoodNet and CNFOOD-241; they are measurements, not quantities derived from the model's own definitions. StyleRM and DCA are presented as explicit modifications of externally published modules (SRM, ref. [50]; criss-cross attention, ref. [51]), and the paper's design equations in Sections 3.3 and 3.4 are standard module definitions rather than a derivation whose conclusion is identical to its input. The ablation in Table 3 compares architectural variants on the same data, and the claimed gains are empirical results, not forced by construction. Several food-specific state-of-the-art methods cited in the introduction (FoodCSwin [4], AlsmViT [31], FoodSwin [32], FGFoodNet [34]) are authored by the same group and are omitted from Table 2, and the baseline training protocol is under-specified; these are legitimate threats to the claim that FE-TResNet 'significantly outperform[s] existing techniques,' but they concern comparison completeness and statistical support, not definitional circularity. Per the review rules, missing baselines and non-load-bearing self-citations do not by themselves constitute circular reasoning, and I found no equation, fit, or citation chain that reduces the paper's central claim to its own inputs.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard supervised-learning assumptions: transfer from ImageNet-21K pretraining, representative public splits, and hand-chosen module placement and hyperparameters. None of these are validated on held-out data or in ablation, so they are ledgered as assumed inputs. The only quantities fitted to data are the network weights themselves, which is ordinary training rather than a derivation. The paper introduces two architectural modules, but no entities of the type this ledger tracks (new particles, forces, dimensions, or conserved quantities); both modules compose existing operations (pooling, convolutions, batch norm, activations, attention maps), so nothing is postulated that needs independent physical evidence.

free parameters (4)
  • Fine-tuning epochs = 100
    Chosen by hand in Table 1; no sensitivity analysis, early stopping, or schedule search is reported.
  • Learning rate and weight decay = 1e-4 and 1e-5
    Chosen by hand in Table 1; no sensitivity analysis is reported, and they directly influence the final accuracies.
  • DCA intermediate channel count C' = unspecified
    Section 3.4 defines C' < C but never states its value; it controls the DCA parameter count and receptive field and is a hidden design choice.
  • AutoAugment policy = unspecified
    Section 4.1 mentions AutoAugment without stating the policy or magnitude; augmentation choices materially affect the reported numbers and the fairness of the comparison.
assumptions (4)
  • domain assumption ImageNet-21K pretrained weights transfer well to Chinese food classification
    Table 1 shows pretraining on ImageNet-21K for 300 epochs with SGD before fine-tuning; the paper gives no from-scratch baseline, so the contribution of pretraining to the reported accuracies is assumed, not measured.
  • domain assumption The public train and test splits of ChineseFoodNet and CNFOOD-241 are representative and the reported accuracies estimate generalization
    Section 3.1 describes the datasets but reports a single split evaluation without cross-validation or confidence bounds.
  • ad hoc to paper Placing StyleRM in basic blocks and DCA in bottleneck blocks is a sound design choice
    Section 3.2 and Figure 2 fix the insertion points; no ablation over positions or theoretical argument justifies them.
  • domain assumption Two shared-parameter DCA passes approximate dense non-local context capture
    Section 3.4 asserts the equivalence based on the Criss-Cross Attention literature; the paper does not measure whether the module actually captures long-range dependencies beyond the small accuracy gain.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Feature-Enhanced TResNet for Fine-Grained Food Image Classification." pith.science (2026). https://pith.science/paper/T7XQ7LCE

@misc{pith2026250712828,
  author       = {Pith},
  title        = {Pith review of: Feature-Enhanced TResNet for Fine-Grained Food Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T7XQ7LCE}},
  note         = {Machine review of arXiv:2507.12828}
}
read the original abstract

Food is not only essential to human health but also serves as a medium for cultural identity and emotional connection. In the context of precision nutrition, accurately identifying and classifying food images is critical for dietary monitoring, nutrient estimation, and personalized health management. However, fine-grained food classification remains challenging due to the subtle visual differences among similar dishes. To address this, we propose Feature-Enhanced TResNet (FE-TResNet), a novel deep learning model designed to improve the accuracy of food image recognition in fine-grained scenarios. Built on the TResNet architecture, FE-TResNet integrates a Style-based Recalibration Module (StyleRM) and Deep Channel-wise Attention (DCA) to enhance feature extraction and emphasize subtle distinctions between food items. Evaluated on two benchmark Chinese food datasets-ChineseFoodNet and CNFOOD-241-FE-TResNet achieved high classification accuracies of 81.37% and 80.29%, respectively. These results demonstrate its effectiveness and highlight its potential as a key enabler for intelligent dietary assessment and personalized recommendations in precision nutrition systems.

Figures

Figures reproduced from arXiv: 2507.12828 by the authors.

Figure 1
Figure 1. These are example images from food dataset. The above two lines are pictures of tofu. Due to different [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall structure of the model. Detailed design of Basicblock and Bottleneck series blocks. The two on the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Detailed design diagram of the StyleRM and SE modules in series. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Detailed design diagram of the SE and DCA modules connected in series. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 41 canonical work pages

  1. [4]

    Foodcswin: A high-accuracy food image recognition model for dietary assessment

    Zhiyong Xiao, Ruke Ling, and Zhaohong Deng. Foodcswin: A high-accuracy food image recognition model for dietary assessment. Journal of Food Composition and Analysis, 139:107110, 2025

  2. [31]

    High accuracy food image classification via vision transformer with data augmentation and feature augmentation

    Xinle Gao, Zhiyong Xiao, and Zhaohong Deng. High accuracy food image classification via vision transformer with data augmentation and feature augmentation. Journal of Food Engineering, 365:111833, 2024

  3. [32]

    Fine grained food image recognition based on swin transformer

    Zhiyong Xiao, Guang Diao, and Zhaohong Deng. Fine grained food image recognition based on swin transformer. Journal of Food Engineering, 380:112134, 2024

  4. [34]

    Fgfoodnet: Ingredient-perceived fine-grained food recognition for dietary monitoring

    Zhiyong Xiao, Yida Sun, and Zhaohong Deng. Fgfoodnet: Ingredient-perceived fine-grained food recognition for dietary monitoring. Journal of Food Measurement and Characterization, 2025 JUN 28 2025

  5. [1]

    Res-vmamba: Fine-grained food category visual classification using selective state space models with deep residual learning

    Chi-Sheng Chen, Guan-Ying Chen, Dong Zhou, Di Jiang, and Dai-Shi Chen. Res-vmamba: Fine-grained food category visual classification using selective state space models with deep residual learning. arXiv preprint arXiv:2402.15761, 2024

  6. [2]

    Deep learning for fine-grained classification of jujube fruit in the natural environment

    Xi Meng, Yingchun Yuan, Guifa Teng, and Tianzhen Liu. Deep learning for fine-grained classification of jujube fruit in the natural environment. Journal of Food Measurement and Characterization, 15(5):4150–4165, 2021

  7. [3]

    Fine-grained food classification methods on the uec food-100 database

    Berker Arslan, Sefer Memi¸ s, Elena Battini Sönmez, and Okan Zafer Batur. Fine-grained food classification methods on the uec food-100 database. IEEE Transactions on Artificial Intelligence, 3(2):238–243, 2021

  8. [5]

    Textural features for image classification

    Robert M Haralick, Karthikeyan Shanmugam, and Its’ Hak Dinstein. Textural features for image classification. IEEE Transactions on systems, man, and cybernetics, (6):610–621, 1973

Show all 61 references
  1. [6]

    On image classification: City images vs

    Aditya Vailaya, Anil Jain, and Hong Jiang Zhang. On image classification: City images vs. landscapes. Pattern recognition, 31(12):1921–1935, 1998

  2. [7]

    Constrained nonnegative matrix factorization and hyperspectral image dimensionality reduction

    Zhiyong Xiao and Salah Bourennane. Constrained nonnegative matrix factorization and hyperspectral image dimensionality reduction. volume 5, pages 46–54, 2014

  3. [8]

    A spectral–spatial similarity-based method and its application to hyperspectral image classification

    Zhiyong) Xiao. A spectral–spatial similarity-based method and its application to hyperspectral image classification. Remote Sensing Letters, pages 122–130, 2014

  4. [9]

    Deep convolutional neural networks for image classification: A comprehensive review

    Waseem Rawat and Zenghui Wang. Deep convolutional neural networks for image classification: A comprehensive review. Neural computation, 29(9):2352–2449, 2017

  5. [10]

    Deep learning for hyperspectral image classification: An overview

    Shutao Li, Weiwei Song, Leyuan Fang, Yushi Chen, Pedram Ghamisi, and Jon Atli Benediktsson. Deep learning for hyperspectral image classification: An overview. IEEE Transactions on Geoscience and Remote Sensing , 57(9):6690–6709, 2019

  6. [11]

    Survey on svm and their application in image classification

    Mayank Arya Chandra and SS Bedi. Survey on svm and their application in image classification. International Journal of Information Technology, 13(5):1–11, 2021

  7. [12]

    Resnet in resnet: Generalizing residual architectures

    Sasha Targ, Diogo Almeida, and Kevin Lyman. Resnet in resnet: Generalizing residual architectures. arXiv preprint arXiv:1603.08029, 2016

  8. [13]

    Wider or deeper: Revisiting the resnet model for visual recognition

    Zifeng Wu, Chunhua Shen, and Anton Van Den Hengel. Wider or deeper: Revisiting the resnet model for visual recognition. Pattern recognition, 90:119–133, 2019

  9. [14]

    Resnet 50

    Brett Koonce and Brett Koonce. Resnet 50. Convolutional neural networks with swift for tensorflow: image recognition and dataset categorization, pages 63–72, 2021. 12 Feature-Enhanced TResNet for Fine-Grained Food Image Classification

  10. [15]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv...

  11. [16]

    Rethinking the inception architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826, 2016

  12. [17]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105–6114. PMLR, 2019

  13. [18]

    Mobilenetv2: Inverted residuals and linear bottlenecks

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510–4520, 2018

  14. [19]

    Searching for mobilenetv3

    Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314–1324, 2019

  15. [20]

    Application of improved convolutional neural network in medical image segmentation

    Chen Liu, Zhiyong Xiao, and Nianmao Du. Application of improved convolutional neural network in medical image segmentation. volume 9, pages 1593–1603, 2019

  16. [21]

    Application of improved convolutional neural network in lung image segmentation

    Baoxin Qian, Zhiyong Xiao, and Wei Song. Application of improved convolutional neural network in lung image segmentation. Journal of Frontiers of Computer Science and Technology, pages 1358–1367, 2020

  17. [22]

    Multi-view hierarchical split network for brain tumor segmentation

    Zhiyong Xiao, Kanghui He, Jianjun Liu, and Weidong Zhang. Multi-view hierarchical split network for brain tumor segmentation. Biomedical Signal Processing and Control, 69:102897, 2021

  18. [23]

    Sr-net: A sequence offset fusion net and refine net for undersampled multislice mr image reconstruction

    Zhiyong Xiao, Nianmao Du, Jianjun Liu, and Weidong Zhang. Sr-net: A sequence offset fusion net and refine net for undersampled multislice mr image reconstruction. Computer Methods and Programs in Biomedicine , 202:105997, 2021

  19. [24]

    Food image segmentation based on deep and shallow dual-branch network

    Zhiyong Xiao, Yang Li, and Zhaohong Deng. Food image segmentation based on deep and shallow dual-branch network. Multimedia Systems, 31:85, 2025

  20. [25]

    3d u-net applied to simple attention module for head and neck tumor segmentation in pet and ct images

    Tao Liu, Yixin Su, Jiabao Zhang, Tianqi Wei, and Zhiyong Xiao. 3d u-net applied to simple attention module for head and neck tumor segmentation in pet and ct images. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Not...

  21. [26]

    Fine-grained crop pest classification based on multi-scale feature fusion and mixed attention mechanisms

    Yiheng Qian, Zhiyong Xiao, and Zhaohong Deng. Fine-grained crop pest classification based on multi-scale feature fusion and mixed attention mechanisms. Frontiers in Plant Science, 16:1500571, 2025

  22. [27]

    Swin attention augmented residual network: a fine-grained pest image recognition method

    Xiang Wang, Zhiyong Xiao, and Zhaohong Deng. Swin attention augmented residual network: a fine-grained pest image recognition method. Frontiers in plant science, page 1619551, 2025

  23. [28]

    Efficient combination of cnn and transformer for dual-teacher uncertainty-guided semi-supervised medical image segmentation

    Zhiyong Xiao, Yixin Su, Zhaohong Deng, and Weidong Zhang. Efficient combination of cnn and transformer for dual-teacher uncertainty-guided semi-supervised medical image segmentation. Computer Methods and Programs in Biomedicine, 226:107099, 2022

  24. [29]

    Rmmlp:rolling mlp and matrix decomposition for skin lesion segmentation

    Chao Ji, Zhaohong Deng, Yan Ding, Fengsheng Zhou, and Zhiyong Xiao. Rmmlp:rolling mlp and matrix decomposition for skin lesion segmentation. Biomedical Signal Processing and Control, 84:104825, 2023

  25. [30]

    Light3dhs: A lightweight 3d hippocampus segmentation method using multiscale convolution attention and vision transformer

    Zhiyong Xiao, Yuhong Zhang, Zhaohong Deng, and Fei Liu. Light3dhs: A lightweight 3d hippocampus segmentation method using multiscale convolution attention and vision transformer. NeuroImage, 292:120608, 2024

  26. [33]

    Highly scalable parallel genetic algorithm on sunway many-core processors

    Zhiyong Xiao, Xu Liu, Jingheng Xu, Qingxiao Sun, and Lin Gan. Highly scalable parallel genetic algorithm on sunway many-core processors. Future Generation Computer Systems, pages 679–691, 2021. 13 Feature-Enhanced TResNet for Fine-Grained Food Image Classification

  27. [35]

    Squeeze-and-excitation networks

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132–7141, 2018

  28. [36]

    Enhance via decoupling: Improving multi-label classifiers with variational feature augmentation

    Ming Chen, Guijin Wang, Jing-Hao Xue, Zijian Ding, and Li Sun. Enhance via decoupling: Improving multi-label classifiers with variational feature augmentation. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1329–1333. Institute of Electrical and Electr...

  29. [37]

    Iml-gcn: Improved multi-label graph convolutional network for efficient yet precise image classification

    Inder Pal Singh, Oyebade Oyedotun, Enjie Ghorbel, and Djamila Aouada. Iml-gcn: Improved multi-label graph convolutional network for efficient yet precise image classification. InAAAI-22 Workshop Program-Deep Learning on Graphs: Methods and Applications, 2022

  30. [40]

    Diagnosis of alzheimer’s disease based on the modified tresnet

    Zelin Xu, Hongmin Deng, Jin Liu, and Yang Yang. Diagnosis of alzheimer’s disease based on the modified tresnet. Electronics, 10(16):1908, 2021

  31. [41]

    Research on x-ray image classification algorithm of covid-19 based on fs-tresn et model

    Yu Zheng. Research on x-ray image classification algorithm of covid-19 based on fs-tresn et model. In 2022 10th International Conference on Information Systems and Computing Technology (ISCTech), pages 598–604. IEEE, 2022

  32. [42]

    Learn from each other to classify better: Cross-layer mutual attention learning for fine-grained visual classification

    Dichao Liu, Longjiao Zhao, Yu Wang, and Jien Kato. Learn from each other to classify better: Cross-layer mutual attention learning for fine-grained visual classification. Pattern Recognition, 140:109550, 2023

  33. [43]

    Transfg: A transformer architecture for fine-grained recognition

    Ju He, Jie-Neng Chen, Shuai Liu, Adam Kortylewski, Cheng Yang, Yutong Bai, and Changhu Wang. Transfg: A transformer architecture for fine-grained recognition. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 852–860, 2022

  34. [44]

    Tresnet: High performance gpu-dedicated architecture

    Tal Ridnik, Hussam Lawen, Asaf Noy, Emanuel Ben Baruch, Gilad Sharir, and Itamar Friedman. Tresnet: High performance gpu-dedicated architecture. In proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1400–1409, 2021

  35. [45]

    Texture synthesis using convolutional neural networks

    Leon Gatys, Alexander S Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. Advances in neural information processing systems, 28, 2015

  36. [46]

    Arbitrary style transfer in real-time with adaptive instance normalization

    Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501–1510, 2017

  37. [47]

    Image style transfer using convolutional neural networks

    Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414–2423, 2016

  38. [48]

    Approximating cnns with bag-of-local-features models works surprisingly well on imagenet

    Wieland Brendel and Matthias Bethge. Approximating cnns with bag-of-local-features models works surprisingly well on imagenet. arXiv preprint arXiv:1904.00760, 2019

  39. [49]

    Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness

    Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018

  40. [50]

    Srm: A style-based recalibration module for convolutional neural networks

    HyunJae Lee, Hyo-Eun Kim, and Hyeonseob Nam. Srm: A style-based recalibration module for convolutional neural networks. In Proceedings of the IEEE/CVF International conference on computer vision, pages 1854–1862, 2019

  41. [51]

    Non-local neural networks

    Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794–7803, 2018

  42. [52]

    Chinesefoodnet: A large-scale image dataset for chinese food recognition

    Xin Chen, Yu Zhu, Hua Zhou, Liang Diao, and Dongyan Wang. Chinesefoodnet: A large-scale image dataset for chinese food recognition. arXiv preprint arXiv:1705.02743, 2017. 14 Feature-Enhanced TResNet for Fine-Grained Food Image Classification

  43. [53]

    Automatic chinese food recognition based on a stacking fusion model

    Bokun Fan, Weiqi Li, Liang Dong, Jingzhen Li, and Zedong Nie. Automatic chinese food recognition based on a stacking fusion model. In 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1–4, 2023

  44. [54]

    Deep networks with stochastic depth

    Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14, pages 646–661. Springer, 2016

  45. [55]

    Improved adam optimizer for deep neural networks

    Zijun Zhang. Improved adam optimizer for deep neural networks. In 2018 IEEE/ACM 26th international symposium on quality of service (IWQoS), pages 1–2. Ieee, 2018

  46. [56]

    On the variance of the adaptive learning rate and beyond

    Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond. arXiv preprint arXiv:1908.03265, 2019

  47. [57]

    Deep classification with linearity-enhanced logits to softmax function

    Hao Shao and Shunfang Wang. Deep classification with linearity-enhanced logits to softmax function. Entropy, 25(5):727, 2023

  48. [58]

    Densenet: Implementing efficient convnet descriptor pyramids

    Forrest Iandola, Matt Moskewicz, Sergey Karayev, Ross Girshick, Trevor Darrell, and Kurt Keutzer. Densenet: Implementing efficient convnet descriptor pyramids. arXiv preprint arXiv:1404.1869, 2014

  49. [59]

    Efficientnetv2: Smaller models and faster training

    Mingxing Tan and Quoc Le. Efficientnetv2: Smaller models and faster training. In International conference on machine learning, pages 10096–10106. PMLR, 2021

  50. [60]

    Inception-v4, inception-resnet and the impact of residual connections on learning

    Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017

  51. [61]

    Xception: Deep learning with depthwise separable convolutions

    François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251–1258, 2017

  52. [62]

    Xception: A technique for the experimental evaluation of dependability in modern computers

    Joao Carreira, Henrique Madeira, and Joao Gabriel Silva. Xception: A technique for the experimental evaluation of dependability in modern computers. IEEE Transactions on Software Engineering, 24(2):125–136, 1998

  53. [63]

    Improved classification of different brain tumors in mri scans using patterned-gridmask

    Ji-hyeon Lee, Jung-woo Chae, and Hyun-chong Cho. Improved classification of different brain tumors in mri scans using patterned-gridmask. IEEE Access, 2024. 15

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.