Pith. sign in

REVIEW 3 major objections 6 minor 38 references

CLPSTNet: A Progressive Multi-Scale Convolutional Steganography Model Integrating Curriculum Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that a progressive multi-scale convolutional block, whose kernel sizes and dilation rates grow stage by stage, lets an image steganography network hide secret bits with higher visual quality and stronger steganalysis…

desk verdict Incremental architecture paper with plausible image-quality tables, but the headline steganalysis-resistance claim is never measured; the security component is asserted, not demonstrated. read the letter →

arxiv 2504.16364 v1 pith:JPNHBTFC submitted 2025-04-23 cs.CV cs.AIcs.CR

Fengchun Liu , Tong Zhang , Chunying Zhang This is my paper · ORCID
classification cs.CVcs.AIcs.CR
keywords imagesteganographyCLPSTNetprogressivemulti-scaleconvolutioncurriculumlearningdilatedInceptionmodulesteganalysisresistancegenerativeadversarialnetwork
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 is trying to establish that a steganography network can hide secret bits inside natural images without visibly degrading them and without leaving traces that a steganalysis network can detect. The proposed CLPSTNet builds both its encoder and decoder from PMCB, a five-branch convolution module that pairs Inception-style multi-kernel pathways with dilated convolutions whose kernel and dilation rate grow progressively through the network. On ALASKA2, VOC2012, and ImageNet, at capacities from 1 to 6 bits per pixel, the authors report PSNR around 50 dB at 1 bpp, SSIM above 0.999, and decoding accuracy of 0.94 to 0.98 at 1 bpp, along with low steganalysis scores for the generated images. If these claims hold, they would mean that curriculum-like progressive receptive-field growth is a practical way to improve both visual imperceptibility and resistance to detection in high-capacity image steganography.

What carries the argument

The load-bearing object is PMCB, the Progressive Multi-scale Convolution Block: five parallel branches, with 1x1, 3x3, and 5x5 convolutions plus two 3x3 dilated convolutions, each large-kernel branch preceded by a 1x1 bottleneck, and all branch outputs concatenated at the end. In CLPSTNet the dilation rates are scheduled progressively, growing from (3,6) to (6,12) to (12,18) through the encoder and decoder, so the effective receptive field expands layer by layer. Dense connections carry shallow edge and detail features forward, and a XuNet-style critic provides the adversarial loss that pushes generated stego images toward low steganalysis scores.

What would settle it

Run CLPSTNet on the ALASKA2 test set at each capacity, extract the recovered 128x128xD bits, and count how many differ from the original secret bits; the count will show whether the reported Accuracy scores certify exact recovery or instead allow a percentage of wrong decoded bits.

Watch

Extended reading notes

Core claim

The central claim is architectural: progressively increasing the receptive field of the embedding and recovery networks, in the same spirit as curriculum learning, is what lets CLPSTNet outperform simpler steganography networks. Each PMCB has five parallel branches, namely 1x1, 3x3, 5x5, and two dilated 3x3 convolutions, and the dilation rates grow from (3,6) to (6,12) to (12,18) across the network, so the shallow layers capture fine local details while deeper layers perceive larger image regions. The authors report that this progressive schedule beats fixed dilation rates, that PMCB plus dense connectivity beats either module alone, and that CLPSTNet surpasses SteganoGAN, HCISNet, and several attention-based baselines on SSIM, MSSSIM, PSNR, and RMSE at 1 bpp on ALASKA2 while keeping decoding accuracy comparable.

Load-bearing premise

The reported decoding success relies on an undefined per-pixel bit-accuracy score; if the intended use is exact recovery of the secret message, a reported 0.98 at 1 bpp does not by itself show that the message comes out intact.

Editorial extensions

If this is right

  • At 1 bpp on all three datasets, CLPSTNet reports PSNR above 49 dB and SSIM above 0.999, which the paper reads as visual near-identity between cover and stego images.
  • Decoding accuracy at 1 bpp is 0.98 on ALASKA2, 0.94 on VOC2012, and 0.95 on ImageNet, but it drops as capacity grows to roughly 0.60 to 0.87 at 3 to 6 bpp, so usable capacity depends on the acceptable error rate.
  • Ablation experiments in the paper show that the PMCB module and the dense connectivity module each contribute to image quality, and that the progressive dilation schedule outperforms fixed dilation rates.
  • In the reported comparison on ALASKA2 at 1 bpp, CLPSTNet beats SteganoGAN, HCISNet, DenseASPP, FC-DenseNet, SENet, ECANet, and CBAM on SSIM, MSSSIM, and PSNR, which is the paper's evidence for its main claim.

Reading between the lines

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

  • In my reading, the 'curriculum' in CLPSTNet is structural rather than data-level: training samples are never sorted by difficulty, so the paper is effectively proposing that curriculum-like architecture can substitute for curriculum scheduling; a natural test is whether adding data-level curriculum on top of the progressive dilation schedule adds further gains.
  • Because the reported decoding accuracy is an average per-pixel score, exact message recovery is not demonstrated; readers who need bit-perfect hidden communication should treat the capacity figures as upper bounds and consider wrapping an error-correcting code around the hidden payload.
  • The steganalysis-resistance claim is tied to the XuNet-style critic used during training; applying a stronger or unseen steganalysis detector to the released model's outputs would show whether the low scores transfer beyond the training adversary.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces CLPSTNet, a convolutional steganography model built from progressive multi-scale convolution blocks (PMCBs) that combine Inception-style branches with dilated convolutions and dense connectivity. The model is trained with an encoder-decoder-critic architecture using a composite loss of embedding quality (MSE, SSIM, MSSSIM), binary cross-entropy for message recovery, and an adversarial steganalysis loss. Experiments on ALASKA2, VOC2012, and ImageNet report PSNR, SSIM, MSSSIM, RMSE, and an 'Accuracy' column for capacities of 1-6 bpp, plus ablations and comparisons with ResNet, SteganoGAN, HCISNet, DenseASPP, and attention-based baselines. The abstract and conclusion claim that the generated stego images have low steganalysis scores, in addition to high visual quality and decoding accuracy.

Significance. If fully supported, the proposed progressive multi-scale architecture could be a useful contribution to deep image steganography, particularly for improving the visual quality of stego images at higher embedding capacities. The paper provides extensive tables and ablations showing large PSNR/SSIM improvements over several baselines, and it ships a public code link. However, the most distinctive claimed advantage—steganalysis resistance—is never measured, and the 'Accuracy' metric is undefined and appears to weaken at high capacities. The manuscript therefore currently establishes an architectural claim about image quality much more convincingly than it establishes the security claim that appears in the abstract.

major comments (3)
  1. [Abstract and Section 5] The abstract and conclusion assert that CLPSTNet generates stego images with 'low steganalysis scores' and achieves 'undetectability of information', yet no steganalysis evaluation appears anywhere in the paper. The Critic described in Section 3.5 is a training-time adversary whose loss is part of the objective; its outputs are never reported, and it is not an independent steganalysis detector. Tables 3-8 contain only image-quality metrics and an undefined Accuracy column. To support the central security claim, the authors must add an experiment with an external steganalysis detector (e.g., detection accuracy, AUC, or false-positive rate on ALASKA2 or another benchmark) comparing CLPSTNet against at least the main baselines; alternatively, the claim must be removed or substantially weakened.
  2. [Section 3.6 and Tables 3-8] The metric labeled 'Accuracy' is never defined. Since the recovery loss in Eq. (9) is binary cross-entropy, Accuracy appears to be per-pixel bit accuracy. Under that interpretation, Table 3 shows roughly 2% of hidden bits are wrong at 1 bpp on ALASKA2, and at 3-6 bpp the accuracy falls to 0.60-0.67, meaning a large fraction of the payload is not recovered bit-exactly. This contradicts the paper's language about 'high decoding accuracy'. The authors must define the metric precisely and, for a steganography claim, report exact-recovery rates or bit-error rates per capacity; otherwise the decoding claim is not interpretable.
  3. [Table 8 and Table 4] The baseline numbers are internally inconsistent. In Table 8, the 'Conv' row reports SSIM=0.98351, MSSSIM=0.99771, PSNR=33.788, RMSE=0.020, Accuracy=0.99, which are exactly the values reported for 'ProgressiveNet' in Table 4, not for the conv-only baseline (Table 4 lists Conv at 1 bpp as SSIM=0.99027, MSSSIM=0.99806, PSNR=34.260, Accuracy=0.74). The same duplicated values appear for ProgressiveNet and Conv in Table 8. This inconsistency undermines the comparison experiment, since the baseline identities are not trustworthy as printed. Please correct the tables and re-run or re-report the comparison.
minor comments (6)
  1. [Section 3.6 and Section 4.2] Equation (5) uses coefficients 'a' and 'b', while the text refers to 'α,β'; please make the notation consistent in both places.
  2. [Section 3.3] The sentence 'Critic chooses the XuNet-based steganalysis auxiliary network proposed by Zhang et al [5]' is duplicated verbatim in the same paragraph; one occurrence should be removed.
  3. [Section 4.1 and Figure 3] The text says 'test results are shown in Figure 3' but later refers to 'Figure 11'; the figure numbering should be corrected.
  4. [Tables 3 and 5] In Table 3, the ALASKA2 row for D=4 has Accuracy 0.66 while D=5 shows 0.67; such non-monotone values are not explained. A brief comment on variance or training details would help readers assess the reliability of the reported numbers.
  5. [Section 4.2] The paper states that the encoding loss coefficients are in the ratio 0.5:0.5:0.3, but Eq. (6) suggests λ1, λ2, λ3 are weights; please clarify whether these are normalized weights or absolute coefficients.
  6. [General] The GitHub link is provided but no license or repository snapshot is cited; consider adding a versioned release or stating the exact commit used for the experiments to support reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the only self-citation is background taxonomy, and the steganalysis-score claim is unsupported rather than derived.

full rationale

The paper's load-bearing comparisons are external: PMCB is evaluated against SteganoGAN, HCISNet, DenseASPP, FC-DenseNet, SENet, ECANet, CBAM, and other baselines on public datasets (ALASKA2, VOC2012, ImageNet), so the reported PSNR/SSIM/decoding-accuracy results do not reduce to the model's own definitions or fitted inputs. The only self-citation, ref [24] in Section 2.2, supplies a background taxonomy of curriculum learning and is not load-bearing for any experimental claim. No uniqueness theorem, ansatz, or fitted parameter is imported from the authors' prior work. The abstract's statement that CLPSTNet images "have low steganalysis scores" is not backed by any reported steganalysis experiment: Tables 3-8 report SSIM, MSSSIM, PSNR, RMSE, and an undefined Accuracy, but never a detection rate or steganalysis score for CLPSTNet or any baseline. That is an evidential gap, not a circular reduction, because the claim is asserted rather than derived from the Critic's loss; the Critic's own output is never reported as evidence. Similarly, "Accuracy" is undefined and appears to be per-pixel bit accuracy under the BCE loss, which is a measurement-definition problem rather than circularity. The conclusion explicitly admits that decoding accuracy degrades with capacity, which further indicates the results are not forced by construction.

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

The central claim depends on several hand-picked hyperparameters (loss weights, dilation schedule) and on domain assumptions about the embedding interface, the critic proxy, and the preprocessing. No new physical or mathematical entities are postulated.

free parameters (4)
  • embedding loss weights (lambda_1, lambda_2, lambda_3) = 0.5, 0.5, 0.3
    Chosen by hand to balance MSE, SSIM and MSSSIM terms in Eq. (6); they affect the reported PSNR/SSIM.
  • total loss coefficients (a, b) = 1, 0.1
    Weights for decoding and steganalysis losses in Eq. (5); the text assigns the steganalysis weight 0.1.
  • dilation rate schedule = [3,6,12,18,3,6,12,18]
    Progressive dilation rates for PMCB modules; this is the key hand-designed architectural choice the paper calls curriculum learning.
  • learning rates and batch size = Adam lr 0.001, SGD lr 0.0001/3, batch 8
    Optimization hyperparameters in Section 4.2; central to reproducing the reported metrics.
assumptions (4)
  • standard math SSIM and MSSSIM definitions (Eqs. 7-8) are standard and taken from prior literature.
    The loss function relies on these similarity metrics without derivation.
  • domain assumption Channel-wise concatenation Cat(Xcover, Ysecret) is a valid interface for embedding arbitrary secret tensors, and the decoder can learn to recover them from the stego image.
    This input representation in Eq. (1) is assumed to be sufficient for steganographic capacity without architectural constraints.
  • domain assumption XuNet-based critic (cited as [5]) produces a valid steganalysis score that can serve as a training signal and a proxy for detectability.
    Section 3.5 uses the critic output s in [0,1]; however, the paper never reports these scores, so the proxy is unvalidated.
  • domain assumption Resizing all datasets to 128x128 via a Matlab program preserves the properties needed for steganography evaluation.
    Section 4.1 states all images were resized; any distortion from downsampling may affect PSNR and steganalysis results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLPSTNet: A Progressive Multi-Scale Convolutional Steganography Model Integrating Curriculum Learning." pith.science (2026). https://pith.science/paper/JPNHBTFC

@misc{pith2026250416364,
  author       = {Pith},
  title        = {Pith review of: CLPSTNet: A Progressive Multi-Scale Convolutional Steganography Model Integrating Curriculum Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JPNHBTFC}},
  note         = {Machine review of arXiv:2504.16364}
}
read the original abstract

In recent years, a large number of works have introduced Convolutional Neural Networks (CNNs) into image steganography, which transform traditional steganography methods such as hand-crafted features and prior knowledge design into steganography methods that neural networks autonomically learn information embedding. However, due to the inherent complexity of digital images, issues of invisibility and security persist when using CNN models for information embedding. In this paper, we propose Curriculum Learning Progressive Steganophy Network (CLPSTNet). The network consists of multiple progressive multi-scale convolutional modules that integrate Inception structures and dilated convolutions. The module contains multiple branching pathways, starting from a smaller convolutional kernel and dilatation rate, extracting the basic, local feature information from the feature map, and gradually expanding to the convolution with a larger convolutional kernel and dilatation rate for perceiving the feature information of a larger receptive field, so as to realize the multi-scale feature extraction from shallow to deep, and from fine to coarse, allowing the shallow secret information features to be refined in different fusion stages. The experimental results show that the proposed CLPSTNet not only has high PSNR , SSIM metrics and decoding accuracy on three large public datasets, ALASKA2, VOC2012 and ImageNet, but also the steganographic images generated by CLPSTNet have low steganalysis scores.You can find our code at \href{https://github.com/chaos-boops/CLPSTNet}{https://github.com/chaos-boops/CLPSTNet}.

Figures

Figures reproduced from arXiv: 2504.16364 by the authors.

Figure 1
Figure 1. Framework of CLPSTNet Feature Maps Concatenate Conv, 1×1 Conv, 1×1 BatchNorm Conv, 1×1 BatchNorm Conv, 3×3 BatchNorm ReLU ReLU ReLU ReLU ReLU Conv, 1×1 Conv, 3×3,dilation=d1 ReLU Conv, 3×3,dilation=d2 Conv, 1×1 BatchNorm Conv, 5×5 BatchNorm BatchNorm BatchNorm BatchNorm ReLU ReLU [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. PMCB Module Fusing the Inception structure and dilation convolution to construct a multi-pathway convolution structure, the Progressive Multi-scale Convolution Block (PMCB) is constructed, starting from a convolution containing a smaller convolution kernel and a dilation rate, to extract the most basic and localized feature information from the feature map. As the network deepens, it gradually expands to a convoluti… view at source ↗
Figure 3
Figure 3. The original image and the stego image generated by CLPSTNet at 1-6 bpp [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 20 canonical work pages

  1. [1]

    Acgis: Adversarial cover generator for image steganography with noise residuals features-preserving

    Junxue Yang and Xin Liao. Acgis: Adversarial cover generator for image steganography with noise residuals features-preserving. Signal Processing: Image Communication, 113:116927, 2023

    2023

  2. [2]

    Channel attention image steganography with generative adversarial networks

    Jingxuan Tan, Xin Liao, Jiate Liu, Yun Cao, and Hongbo Jiang. Channel attention image steganography with generative adversarial networks. IEEE transactions on network science and engineering, 9(2):888–903, 2021

    work page 2021

  3. [3]

    Generating steganographic images via adversarial training

    Jamie Hayes and George Danezis. Generating steganographic images via adversarial training. Advances in neural information processing systems, 30, 2017

    2017

  4. [4]

    Steganogan: High capacity image steganography with gans

    Kevin Alex Zhang, Alfredo Cuesta-Infante, Lei Xu, and Kalyan Veeramachaneni. Steganogan: High capacity image steganography with gans. arXiv preprint arXiv:1901.03892, 2019

    arXiv 1901

  5. [5]

    Invisible steganography via generative adversarial networks

    Ru Zhang, Shiqi Dong, and Jianyi Liu. Invisible steganography via generative adversarial networks. Multimedia tools and applications, 78(7):8559–8575, 2019

    2019

  6. [6]

    High-capacity image steganography based on improved fc-densenet

    Xintao Duan, Liu Nao, Gou Mengxiao, Dongli Yue, Zimei Xie, Yuanyuan Ma, and Chuan Qin. High-capacity image steganography based on improved fc-densenet. IEEE Access, 8:170174–170182, 2020

    work page 2020

  7. [7]

    Hinet: Deep image hiding by invertible network

    Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4733–4742, 2021

    2021

  8. [8]

    Hidinggan: High capacity information hiding with generative adversarial network

    Zihan Wang, Neng Gao, Xin Wang, Ji Xiang, Daren Zha, and Linghui Li. Hidinggan: High capacity information hiding with generative adversarial network. In Computer Graphics Forum, volume 38, pages 393–401. Wiley Online Library, 2019

    work page 2019

Show all 38 references
  1. [9]

    Progressive growing of gans for improved quality, stability, and variation

    Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017

    2017 arXiv

  2. [10]

    Curriculum dropout

    Pietro Morerio, Jacopo Cavazza, Riccardo V olpi, René Vidal, and Vittorio Murino. Curriculum dropout. In Proceedings of the IEEE international conference on computer vision, pages 3544–3552, 2017

    2017

  3. [11]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020

    2020

  4. [12]

    Advanced image steganog- raphy using a u-net-based architecture with multi-scale fusion and perceptual loss

    Lu Zeng, Ning Yang, Xiang Li, Aidong Chen, Hongyuan Jing, and Jiancheng Zhang. Advanced image steganog- raphy using a u-net-based architecture with multi-scale fusion and perceptual loss. Electronics, 12(18):3808, 2023

    2023

  5. [13]

    Joint adjustment image steganography networks

    Le Zhang, Yao Lu, Tong Li, and Guangming Lu. Joint adjustment image steganography networks. Signal Processing: Image Communication, 118:117022, 2023

    2023

  6. [14]

    Image steganography with deep orthogonal fusion of multi-scale channel attention

    Yinyin Peng, Donghui Hu, Gang Pei, and Yaofei Wang. Image steganography with deep orthogonal fusion of multi-scale channel attention. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4850–4854. IEEE, 2024

    2024

  7. [15]

    High invisibility image steganography with wavelet transform and generative adversarial network

    Ye Yao, Junyu Wang, Qi Chang, Yizhi Ren, and Weizhi Meng. High invisibility image steganography with wavelet transform and generative adversarial network. Expert Systems with Applications, 249:123540, 2024

    2024

  8. [16]

    Enhanced pixel privacy: Leverag- ing deep learning for advanced image steganography

    Mohammmad Kashif, Iman Shakeel, Shahnawaz Ahmad, and Shabana Mehfuz. Enhanced pixel privacy: Leverag- ing deep learning for advanced image steganography. In 2023 7th International Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud)(I-SMAC), pages 535–541. IEEE, 2023. 14

    2023

  9. [17]

    Adversarial feature hybrid framework for steganography with shifted window local loss

    Zhengze Li, Xiaoyuan Yang, Kangqing Shen, Fazhen Jiang, Jin Jiang, Huwei Ren, and Yixiao Li. Adversarial feature hybrid framework for steganography with shifted window local loss. Neural Networks, 165:358–369, 2023

    2023

  10. [18]

    Stegformer: rebuilding the glory of autoencoder-based steganography

    Xiao Ke, Huanqi Wu, and Wenzhong Guo. Stegformer: rebuilding the glory of autoencoder-based steganography. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 2723–2731, 2024

    2024

  11. [19]

    Robust invertible image steganography

    Youmin Xu, Chong Mou, Yujie Hu, Jingfen Xie, and Jian Zhang. Robust invertible image steganography. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7875–7884, 2022

    2022

  12. [20]

    Image hide with invertible network and swin transformer

    Yuhuan Feng, Yunjie Liu, Hongjuan Wang, Jin Dong, Rujia Wang, and Chunpeng Tian. Image hide with invertible network and swin transformer. In International Conference on Data Mining and Big Data , pages 385–394. Springer, 2022

    2022

  13. [21]

    iscmis: Spatial-channel attention based deep invertible network for multi-image steganography

    Fengyong Li, Yang Sheng, Xinpeng Zhang, and Chuan Qin. iscmis: Spatial-channel attention based deep invertible network for multi-image steganography. IEEE Transactions on Multimedia, 26:3137–3152, 2023

    2023

  14. [22]

    Deepmih: Deep invertible network for multiple image hiding

    Zhenyu Guan, Junpeng Jing, Xin Deng, Mai Xu, Lai Jiang, Zhou Zhang, and Yipeng Li. Deepmih: Deep invertible network for multiple image hiding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):372–390, 2022

    2022

  15. [23]

    Curriculum learning

    Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41–48, 2009

    2009

  16. [24]

    A review of the evaluation system for curriculum learning

    Fengchun Liu, Tong Zhang, Chunying Zhang, Lu Liu, Liya Wang, and Bin Liu. A review of the evaluation system for curriculum learning. Electronics, 12(7):1676, 2023

    2023

  17. [25]

    Attention-guided curriculum learning for weakly supervised classification and localization of thoracic diseases on chest radiographs

    Yuxing Tang, Xiaosong Wang, Adam P Harrison, Le Lu, Jing Xiao, and Ronald M Summers. Attention-guided curriculum learning for weakly supervised classification and localization of thoracic diseases on chest radiographs. In Machine Learning in Medical Imaging: 9th International ...

    2018

  18. [26]

    Cased: curriculum adaptive sampling for extreme data imbalance

    Andrew Jesson, Nicolas Guizard, Sina Hamidi Ghalehjegh, Damien Goblot, Florian Soudan, and Nicolas Chapados. Cased: curriculum adaptive sampling for extreme data imbalance. In International conference on medical image computing and computer-assisted intervention, pages 639–646...

    2017

  19. [27]

    Acpl: Anti- curriculum pseudo-labelling for semi-supervised medical image classification

    Fengbei Liu, Yu Tian, Yuanhong Chen, Yuyuan Liu, Vasileios Belagiannis, and Gustavo Carneiro. Acpl: Anti- curriculum pseudo-labelling for semi-supervised medical image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2...

    2022

  20. [28]

    A curriculum domain adaptation approach to the semantic segmentation of urban scenes

    Yang Zhang, Philip David, Hassan Foroosh, and Boqing Gong. A curriculum domain adaptation approach to the semantic segmentation of urban scenes. IEEE transactions on pattern analysis and machine intelligence , 42(8):1823–1841, 2019

    2019

  21. [29]

    Curriculum learning of multiple tasks

    Anastasia Pentina, Viktoriia Sharmanska, and Christoph H Lampert. Curriculum learning of multiple tasks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5492–5500, 2015

    2015

  22. [30]

    Curriculum by smoothing

    Samarth Sinha, Animesh Garg, and Hugo Larochelle. Curriculum by smoothing. Advances in Neural Information Processing Systems, 33:21653–21664, 2020

    2020

  23. [31]

    Curriculum based dropout discriminator for domain adaptation

    Vinod Kumar Kurmi, Vipul Bajaj, Venkatesh K Subramanian, and Vinay P Namboodiri. Curriculum based dropout discriminator for domain adaptation. arXiv preprint arXiv:1907.10628, 2019

    1907 arXiv

  24. [32]

    Improved training with curriculum gans

    Rishi Sharma, Shane Barratt, Stefano Ermon, and Vijay Pande. Improved training with curriculum gans. arXiv preprint arXiv:1807.09295, 2018

    2018 arXiv

  25. [33]

    Going deeper with convolutions

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1–9, 2015

    2015

  26. [34]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  27. [35]

    Denseaspp for semantic segmentation in street scenes

    Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic segmentation in street scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3684–3692, 2018

    2018

  28. [36]

    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

    2018

  29. [37]

    Eca-net: Efficient channel attention for deep convolutional neural networks

    Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11534–11542, 2020. 15

    2020

  30. [38]

    Cbam: Convolutional block attention module

    Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 16

    2018

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.