Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Enhancing Frequency for Single Image Super-Resolution with Learnable Separable Kernels

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Replacing square kernels with learnable rank-one ones cuts super-resolution cost and sharpens high-frequency detail.

desk verdict The separable-kernel trick is a known low-rank factorization with clean complexity math, but the frequency story is unproven and the abstract overstates the gains; the scale-dependent trend is fragile but worth one careful look. read the letter →

arxiv 2506.04555 v2 pith:ISMDHVCR submitted 2025-06-05 cs.CV cs.MM

Heng Tian This is my paper
classification cs.CVcs.MM
keywords singleimagesuper-resolutionlearnableseparablekernelsrank-onematrixfrequencyenhancementparameterreductionefficientconvolution
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

This paper proposes replacing the usual square convolution kernels in single-image super-resolution networks with Learnable Separable Kernels (LSKs), which are sums of rank-one matrices, each decomposable into two orthogonal one-dimensional kernels. The authors claim that because rank-one separable kernels can act as high-frequency enhancers, their use directly improves the reconstruction of lost detail while cutting parameters and computation by more than 60%. The paper reports that separable versions of SRCNN, ESPCN, and VDSR match or slightly beat the original models on five benchmarks, and that the relative advantage grows as the upscaling factor increases. If correct, this offers a plug-and-play way to make super-resolution models lighter and better at recovering high-frequency content without auxiliary loss functions.

What carries the argument

The central object is the Learnable Separable Kernel (LSK), an n×n matrix of rank one that factors as a product of an n×1 and a 1×n vector. The paper replaces a square convolution layer with two consecutive 1D convolution layers without an activation between them, so the two operations can be merged back into a single separable kernel at inference. Because the intermediate feature map is computed without nonlinearity, the whole two-layer operation is algebraically equivalent to a sum of LSKs, and this decomposition is what produces the parameter and FLOP reductions while preserving the representational capacity of a full-rank kernel. The frequency-enhancement argument rests on the observation that rank-one high-pass operators such as Sobel are separable, so the learned LSKs are expected to behave as frequency enhancers during super-resolution.

What would settle it

Train S-SRCNN on a standard benchmark, extract the learned 1D kernels from an LSK layer, form their outer products, and compute the 2D Fourier transform of each kernel; if the average response at high spatial frequencies is not greater than the response at low frequencies, the claimed frequency-enhancement mechanism is contradicted.

Watch

Extended reading notes

Core claim

The central claim is that a square convolution kernel in a super-resolution network can be replaced by a sum of rank-one separable kernels, called LSKs, and that this replacement does two things at once: it cuts the parameter count and FLOPs to about 2/n_k of the original for an n_k×n_k kernel, and it gives the network an inherent tendency to enhance image frequency components. The paper reasons that edge-detecting kernels such as Sobel are separable and high-pass, so learning rank-one kernels should let the network acquire frequency-enhancing filters. Empirically, S-SRCNN, S-ESPCN, and S-VDSR (the separable versions) reduce parameters and FLOPs by over 60% in the SRCNN case and 32–43% in the VDSR/ESPCN cases, while staying within about 0.1 dB of the original models in PSNR. The paper also reports that on ×4 upscaling, the separable versions win more often than the normal versions on both PSNR and SSIM, consistent with the idea that larger scale factors lose more high-frequency information and therefore benefit more from frequency-enhancing kernels.

Load-bearing premise

The whole frequency-enhancement argument rests on the premise that a learned rank-one kernel will naturally behave as a high-frequency enhancer, but rank-one kernels can just as easily learn low-pass or smoothing filters, and the paper provides no spectral measurement of the learned weights to rule that out.

Editorial extensions

If this is right

  • Models using LSKs cut parameters and FLOPs by roughly 2/n_k, so a 5×5 kernel replacement halves the cost and a 3×3 replacement reduces it to two-thirds, yielding the reported >60% reduction in the SRCNN case.
  • LSK-based super-resolution is compatible with pre-upsampling, post-upsampling, and residual-based architectures, so the module is genuinely plug-and-play across the main SISR framework families.
  • The relative performance of LSK models improves as the scale factor goes from ×2 to ×4, implying that the benefit grows precisely when the low-resolution input has lost more high-frequency content.
  • Because the two 1D kernels are merged during inference, the computational savings are realized at deployment time, not only during training.
  • LSK models produce feature maps that show less separation between frequency-related and detail-lacking channels, which the paper interprets as the kernels integrating frequency enhancement into each feature rather than isolating it.

Reading between the lines

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

  • The paper does not measure the frequency response of the learned LSK weights, so the claimed mechanism — that learned rank-one kernels will be high-pass — remains unverified; a test that computes the FFT of trained LSKs would directly probe this assumption.
  • Because a sum of n_fe rank-one matrices has rank at most n_fe, the expressiveness of the replaced square layer is bounded by the number of intermediate feature maps, so the method implicitly trades a controlled amount of capacity for efficiency.
  • The observed improvement at larger scale factors may partly reflect a regularization effect of the lower-parameter separable architecture, which could help on harder tasks where the baseline overfits; this alternative explanation is not ruled out by the experiments.
  • The authors suggest combining LSKs with auxiliary structures; a natural extension is pairing LSKs with a frequency-aware loss, such as gradient variance loss, to test whether the two frequency-enhancement mechanisms are additive.
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 paper proposes Learnable Separable Kernels (LSKs), rank-one convolutional kernels implemented as two 1D convolutions, as a plug-and-play replacement for square kernels in single image super-resolution. The authors argue from a frequency perspective that LSKs enhance high-frequency content, derive parameter and FLOP reductions, and evaluate the replacement on SRCNN, ESPCN, and VDSR variants across five benchmark datasets. They report large parameter/computation savings and comparable or improved PSNR/SSIM, with larger relative gains at higher upscaling factors.

Significance. If the frequency-enhancement claim were established, the paper would offer a simple and broadly applicable low-rank reparameterization for SISR, with exact complexity reductions that are verified in Table 3. The complexity analysis in Sec. 3.3 is arithmetically correct, and the experimental comparison across three architectures and five datasets is a useful empirical study of kernel separability. The main weakness is that the mechanistic explanation in terms of frequency enhancement is not supported by the evidence provided, and the abstract overstates the parameter savings. The contribution is therefore currently a sound efficiency technique with an unsubstantiated frequency-based motivation.

major comments (3)
  1. [Sec. 3.2 (and Sec. 3.1, final paragraph)] The central frequency claim is not established. The paper argues that LSKs enhance frequency because Sobel, a separable rank-one kernel, is high-pass. However, rank-one kernels can equally be low-pass: the 3x3 box filter is the outer product of two [1,1,1] vectors, and the learned 1D kernels are unconstrained. The manuscript provides no Fourier-domain analysis of learned kernels or feature maps, and Sec. 4.4's two-example visualization is qualitative. A direct spectral measurement of trained LSKs or of output feature maps, or a constrained formulation that enforces high-pass behavior, is needed before the frequency mechanism can be accepted.
  2. [Abstract and Table 3] The abstract's claim of 'over 60% reduction in both the number of parameters and computational requirements' as a general property of the baseline methods is contradicted by Table 3: only S-SRCNN exceeds 60% (62.45%), while S-ESPCN is 43.14% and the S-VDSR variants are 32–35%. The sentence should be restricted to the SRCNN case or rephrased to report per-method reductions.
  3. [Sec. 4.3 and Table 4] The scale-dependent benefit is not convincingly established. Table 4 shows that at x2 the separable variants win only 33% of PSNR comparisons and 10% of SSIM comparisons, and many differences in Table 2 are below 0.1 dB with no error bars or significance tests. The monotone trend in Fig. 3 is consistent with the frequency hypothesis, but it could also reflect optimization or parameterization effects; the post hoc explanation does not constitute a predictive test.
minor comments (5)
  1. [Sec. 3.1, Eq. (3)] The product of an n x 1 and a 1 x n kernel is an outer product, not a convolution, and the '*' notation is used ambiguously; this should be clarified.
  2. [Abstract and Sec. 5] The word 'orthogonal' in describing the 1D kernels is not defined, and no orthogonality constraint is imposed on the learned kernels.
  3. [Sec. 4.3, Table 4] The proportions in Table 4 are not accompanied by the number of comparisons or confidence intervals, which makes values such as 0.33 versus 0.67 difficult to interpret.
  4. [Table 3 caption] The caption states that FLOPs are calculated when upscaling to 512x512 on x2, x3, and x4 tasks, but the text does not clarify whether the same input resolution is used for all scales; this should be stated explicitly.
  5. [Fig. 3] The axis labels in Fig. 3 are not legible in the provided version, making the trend difficult to verify.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the rank-one frequency claim is an invalid generalization, not a definitional equivalence or self-citation chain.

full rationale

Walking the derivation chain: LSKs are defined as rank-one matrices (Sec. 3.1), and the replacement of square kernels by sums of such matrices is given by Eqs. (3)-(4). The load-bearing frequency claim appears in Sec. 3.2: "A rank-1 matrix can enhance the frequency," supported only by the Sobel example. This is an overgeneralization, since a rank-one outer product can also be a box blur or other low-pass filter, and the paper offers no spectral analysis of the learned kernels. However, this is not circular: the definition of LSK does not build frequency enhancement into the term, the assertion is not derived by substituting equations into one another, and no fitted parameter is renamed as a prediction. The scale-dependent performance trend in Sec. 4.3 and Table 4 is an empirical observation made after the experiments; it is consistent with the hypothesis but was not an output forced by construction. The parameter and FLOP reductions in Sec. 3.3 are direct analytical calculations, independently tabulated in Table 3; the abstract's "over 60%" phrasing overstates the reductions for ESPCN and VDSR, but that is a reporting overclaim rather than a circular step. There are no load-bearing self-citations or imported uniqueness theorems. The central weakness, namely that rank-one does not by itself imply high-frequency enhancement, is a correctness/evidence gap that would require direct spectral measurement, not a circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or architectural entities of the graviton type; its contribution is a factorization of existing convolution operations. The main uncharged assumptions are the frequency-enhancement premise and the equal-channel-count simplification.

free parameters (2)
  • n_fe (number of intermediate feature maps in the separable layer) = n_fe = n_fm (set equal to the output feature map count of the next layer)
    Hand-chosen design choice stated in Sec. 4.1; no ablation over this value.
  • Layers selected for replacement = Only layers where n_fm-1 > 1 and n_fm > 1 (e.g., layer 2 of SRCNN)
    Chosen to avoid parameter increases at input/output layers; this selection largely determines the reported reduction percentages.
assumptions (3)
  • standard math Convolution is linear, allowing two 1D convolutions without an intermediate activation to be merged into a single convolution.
    Used in Eq. (3) to show the equivalence between the two-layer 1D decomposition and a sum of rank-one kernels.
  • ad hoc to paper Rank-one kernels enhance image frequency content, as exemplified by Sobel.
    Stated in Sec. 3.2 as the core justification for LSKs; no Fourier analysis or spectral measurement is provided.
  • domain assumption The complexity analysis assumes n_fm-1 = n_fe = n_fm to derive simple ratios.
    Used in Sec. 3.3 (Eqs. 7 and 13); the actual networks do not satisfy this assumption (e.g., SRCNN layer 2 has 64 input and 32 output channels), so the ratios are approximate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Frequency for Single Image Super-Resolution with Learnable Separable Kernels." pith.science (2026). https://pith.science/paper/ISMDHVCR

@misc{pith2026250604555,
  author       = {Pith},
  title        = {Pith review of: Enhancing Frequency for Single Image Super-Resolution with Learnable Separable Kernels},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ISMDHVCR}},
  note         = {Machine review of arXiv:2506.04555}
}
read the original abstract

Existing approaches often enhance the performance of single-image super-resolution (SISR) methods by incorporating auxiliary structures, such as specialized loss functions, to indirectly boost the quality of low-resolution images. In this paper, we propose a plug-and-play module called Learnable Separable Kernels (LSKs), which are formally rank-one matrices designed to directly enhance image frequency components. We begin by explaining why LSKs are particularly suitable for SISR tasks from a frequency perspective. Baseline methods incorporating LSKs demonstrate a significant reduction of over 60\% in both the number of parameters and computational requirements. This reduction is achieved through the decomposition of LSKs into orthogonal and mergeable one-dimensional kernels. Additionally, we perform an interpretable analysis of the feature maps generated by LSKs. Visualization results reveal the capability of LSKs to enhance image frequency components effectively. Extensive experiments show that incorporating LSKs not only reduces the number of parameters and computational load but also improves overall model performance. Moreover, these experiments demonstrate that models utilizing LSKs exhibit superior performance, particularly as the upscaling factor increases.

Figures

Figures reproduced from arXiv: 2506.04555 by the authors.

Figure 1
Figure 1. Lsks can be decomposed into two 1D kernels, which improves the efficiency and reduces the number of parameters. It is worth noting that LSKs are different from Inception-v3 [18], which replaces any n×n convolutions by a 1×n convolutions followed by a n × 1 convolutions. – Based on the specificity of the SISR task, we introduce the LSKs into the SISR task to enhance frequency combined with the assistance of neural ne… view at source ↗
Figure 2
Figure 2. The process of replacing the square kernels with LSKs and decomposing LSKs into 1D kernels. There will be an extra layer of feature maps, since the square kernels become two-layer 1D kernels. 3.1 Learnable separable kernel LSKs are essentially rank-one n×n learnable matrices, which can be decomposed into two multiplicative n × 1 and 1 × n matrices. As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The line graph of the proportion of better versions under different scale factors in PSNR and SSIM. The trend of the separable version methods is gradually improving with the increase of scale factor. It is worth noting that although the normal version sometimes outperforms the separable version in terms of PSNR, this advantage does not exceed 0.1 dB [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The comparison of the last layer feature maps between the SRCNN and S￾SRCNN with scale factor ×4. Feature maps related to frequency and lacking details are marked with red boxes and green boxes, separately. In the last layer of feature maps of the SRCNN, we can see tha…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 24 canonical work pages

  1. [1]

    In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Abrahamyan, L., Truong, A.M., Philips, W., Deligiannis, N.: Gradient variance loss for structure-enhanced image super-resolution. In: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 3219–3223. IEEE (2022)

    work page 2022

  2. [2]

    ACM Journal on Emerging Technologies in Computing Systems (JETC) 13(3), 1–18 (2017)

    Anwar, S., Hwang, K., Sung, W.: Structured pruning of deep convolutional neural networks. ACM Journal on Emerging Technologies in Computing Systems (JETC) 13(3), 1–18 (2017)

    work page 2017

  3. [3]

    In: Pro- ceedings of the British Machine Vision Conference

    Bevilacqua, M., Roumy, A., Guillemot, C., line Alberi Morel, M.: Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In: Pro- ceedings of the British Machine Vision Conference. pp. 135.1–135.10. BMVA Press (2012)

    work page 2012

  4. [4]

    In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV)

    Ding, X., Guo, Y., Ding, G., Han, J.: Acnet: Strengthening the kernel skeletons for powerful cnn via asymmetric convolution blocks. In: 2019 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 1911–1920 (2019)

    work page 2019

  5. [5]

    In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Ding, X., Zhang, X., Han, J., Ding, G.: Diverse branch block: Building a convolu- tion as an inception-like unit. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10881–10890 (2021)

    work page 2021

  6. [6]

    In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: Repvgg: Making vgg-style convnets great again. In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13728–13737 (2021)

    work page 2021

  7. [7]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 38(2), 295–307 (2016)

    Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolu- tional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 38(2), 295–307 (2016)

    work page 2016

  8. [8]

    http://www.pascal-network.org/challenges/VOC/voc2012/workshop/index.html

    Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascal-network.org/challenges/VOC/voc2012/workshop/index.html

    2012

Show all 29 references
  1. [9]

    He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016)

    2016

  2. [10]

    In: 2015 IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR)

    Huang, J.B., Singh, A., Ahuja, N.: Single image super-resolution from transformed self-exemplars. In: 2015 IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR). pp. 5197–5206 (2015)

    2015

  3. [11]

    In: Proceedings of the British Machine Vision Conference (2014)

    Jaderberg, M., Vedaldi, A., Zisserman, A.: Speeding up convolutional neural net- works with low rank expansions. In: Proceedings of the British Machine Vision Conference (2014)

    2014

  4. [12]

    Kim, J., Lee, J.K., Lee, K.M.: Accurate image super-resolution using very deep convolutionalnetworks.In:2016IEEEConferenceonComputerVisionandPattern Recognition (CVPR). pp. 1646–1654 (2016)

    2016

  5. [13]

    Irbm42(2), 120–133 (2021)

    Li, Y., Sixou, B., Peyrin, F.: A review of the deep learning methods for medical images super resolution problems. Irbm42(2), 120–133 (2021)

    2021

  6. [14]

    In: Proceedings Eighth IEEE International Conference on Computer Vision

    Martin, D., Fowlkes, C., Tal, D., Malik, J.: A database of human segmented natu- ral images and its application to evaluating segmentation algorithms and measur- ing ecological statistics. In: Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001. vol....

    2001

  7. [15]

    Multimedia Tools and Applications76, 21811–21838 (2017) Title Suppressed Due to Excessive Length 15

    Matsui, Y., Ito, K., Aramaki, Y., Fujimoto, A., Ogawa, T., Yamasaki, T., Aizawa, K.: Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications76, 21811–21838 (2017) Title Suppressed Due to Excessive Length 15

    2017

  8. [16]

    In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Shi, W., Caballero, J., Huszár, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVP...

    2016

  9. [17]

    Neurocomputing487, 46–65 (2022)

    Su, J., Xu, B., Yin, H.: A survey of deep learning approaches to image restoration. Neurocomputing487, 46–65 (2022)

    2022

  10. [18]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the incep- tion architecture for computer vision. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2818–2826 (2016)

    2016

  11. [19]

    In: 2013 IEEE International Conference on Com- puter Vision

    Timofte, R., De, V., Gool, L.V.: Anchored neighborhood regression for fast example-based super-resolution. In: 2013 IEEE International Conference on Com- puter Vision. pp. 1920–1927 (2013)

    2013

  12. [20]

    In: Computer Vision–ACCV 2014: 12th Asian Conference on Computer Vision, Singapore, Singapore, November 1-5, 2014, Re- vised Selected Papers, Part IV 12

    Timofte, R., De Smet, V., Van Gool, L.: A+: Adjusted anchored neighborhood regression for fast super-resolution. In: Computer Vision–ACCV 2014: 12th Asian Conference on Computer Vision, Singapore, Singapore, November 1-5, 2014, Re- vised Selected Papers, Part IV 12. pp. 111–12...

    2015

  13. [21]

    In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

    Wang, Y.: Edge-enhanced feature distillation network for efficient super-resolution. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 776–784 (2022)

    2022

  14. [22]

    IEEE Transactions on Pattern Analysis and Machine Intelligence43(10), 3365– 3387 (2021)

    Wang,Z.,Chen,J.,Hoi,S.C.H.:Deeplearningforimagesuper-resolution:Asurvey. IEEE Transactions on Pattern Analysis and Machine Intelligence43(10), 3365– 3387 (2021)

    2021

  15. [23]

    IEEE Transactions on Image Processing 13(4), 600–612 (2004)

    Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600–612 (2004)

    2004

  16. [24]

    In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV)

    Xie, W., Song, D., Xu, C., Xu, C., Zhang, H., Wang, Y.: Learning frequency-aware dynamic network for efficient super-resolution. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4288–4297 (2021)

    2021

  17. [25]

    IEEE Transactions on Image Processing19(11), 2861–2873 (2010)

    Yang, J., Wright, J., Huang, T.S., Ma, Y.: Image super-resolution via sparse rep- resentation. IEEE Transactions on Image Processing19(11), 2861–2873 (2010)

    2010

  18. [26]

    In: Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7

    Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse- representations. In: Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7. pp. 711–730. Springer (2012)

    2012

  19. [27]

    In: 2021 IEEE/CVF Inter- national Conference on Computer Vision (ICCV)

    Zhan, Z., Gong, Y., Zhao, P., Yuan, G., Niu, W., Wu, Y., Zhang, T., Jayaweera, M., Kaeli, D., Ren, B., Lin, X., Wang, Y.: Achieving on-mobile real-time super- resolution with neural architecture and pruning search. In: 2021 IEEE/CVF Inter- national Conference on Computer Visio...

    2021

  20. [28]

    IEEE Transactions on Pattern Analysis and Ma- chine Intelligence38(10), 1943–1955 (2016)

    Zhang, X., Zou, J., He, K., Sun, J.: Accelerating very deep convolutional networks for classification and detection. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence38(10), 1943–1955 (2016)

    2016

  21. [29]

    In: Proceedings of the 29th ACM International Conference on Multimedia

    Zhang, X., Zeng, H., Zhang, L.: Edge-oriented convolution block for real-time su- per resolution on mobile devices. In: Proceedings of the 29th ACM International Conference on Multimedia. pp. 4034–4043 (2021)

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.