Deep Evolution for Facial Emotion Recognition

Bruce A. Bassett; Emmanuel Dufourq

REVIEW 4 major objections 5 minor 41 references

Deep Evolution for Facial Emotion Recognition

T0 review · 4 major / 5 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read EvoFER's evolutionary search finds small nose-relative face patches that let a fixed CNN match or beat full-face training while cutting trainable parameters by about 95%.

desk verdict A novel evolutionary patch-selection idea for FER that is currently under-supported because the reported gains are not separated from the trivial effect of using smaller inputs; worth refereeing if the authors add a random-patch control. read the letter →

arxiv 2009.14194 v2 pith:2ZORRICS submitted 2020-09-29 cs.NE cs.CVcs.LG

Emmanuel Dufourq , Bruce A. Bassett This is my paper · ORCID

classification cs.NEcs.CVcs.LG

keywords evolutionaryalgorithmfacialexpressionrecognitionconvolutionalneuralnetworkpatchextractionparameterreductioncompressionspatialattentioninterpretability

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

EvoFER, the paper's evolutionary algorithm, is built around an odd bet: a convolutional network trained for facial emotion recognition does not need the whole face. The algorithm encodes candidate attention patterns as chromosomes—patch widths and heights plus nose-relative coordinates—and searches for a small set of patches whose stacked image, fed into a fixed small CNN, preserves or improves classification accuracy while shrinking the input. On four expression datasets the best evolved patch sets cut trainable parameters by 88% to 98% (about 95% on average) while the same CNN trained on those patches beat the same CNN trained on full images in every dataset. If true, the idea turns input selection into a form of learned spatial attention that is cheap, inspectable, and independent of the classifier architecture.

What carries the argument

The load-bearing object is the EvoFER chromosome: two genes encode patch width $\alpha$ and height $\beta$, and each additional $(x,y)$ pair encodes the top-left corner of one patch relative to the automatically detected nose. A chromosome is applied to every training image, the patches are extracted and stacked vertically to form a new multi-channel image, and that image replaces the full face as input to an otherwise fixed CNN. The multi-objective fitness function $$\text{Fitness} = \exp\left\{ W_S \left(\frac{S_c}{S_b}\right) + \frac{P_b - P_c}{P_b} \right\}$$ balances validation accuracy $S_c/S_b$ against parameter savings $(P_b-P_c)/P_b$, with $W_S$ a user-set weight; mutation adds, removes, changes, or locally shifts patches, and crossover swaps patch coordinates between parents. Because the same CNN and hyperparameters are used for baseline and EvoFER runs, any performance difference comes from the input patches rather than the architecture.

What would settle it

Train the baseline full-image CNN on JAFFE and MUG for enough epochs that validation accuracy plateaus, and rerun the EvoFER comparison with identical test splits; if the converged full-image network matches or exceeds EvoFER's patch accuracy, the claim that the evolved patches retain or improve accuracy is not general.

Watch

Extended reading notes

Core claim

The paper's central claim is that a fixed CNN trained on a small stacked image of nose-relative face patches can match or beat itself trained on the full face, while using on average 95% fewer trainable parameters. Concretely, in ten repeated runs on KDEF, JAFFE, RAFD, and MUG, EvoFER raised test accuracy over the baseline CNN by 1.3, 15.5, 5.0, and 20.6 percentage points respectively, with parameter reductions of 88%, 97.9%, 97.9%, and 98.2%. The evolved patches converge on the eyes, eyebrows, and mouth regions, which the paper reads as evidence that the method implements a novel form of attention and that the most discriminative expression information is local. The authors state the result in the abstract as a robust average reduction of 95% in parameters with no loss in classification accuracy.

Load-bearing premise

The comparison assumes that a CNN trained on full images for ten epochs is a fair reference, so if that baseline is not converged, EvoFER's accuracy gains are inflated by convergence speed rather than by the patches themselves.

Editorial extensions

If this is right

On the paper's evidence, any fixed CNN can be made roughly 20 times smaller on the input side by training it on evolved stacked patches rather than full faces, with the same architecture and hyperparameters.
Because the evolved patches consistently land on eyes and mouth, the method doubles as a tool for inspecting what a face-expression classifier relies on; one can literally see the regions driving predictions.
Training time for the fixed CNN drops sharply once the patches are found (for example, KDEF drops from 247 to 42 seconds in the paper's reported per-dataset training time), making the approach usable on modest hardware.
The fitness weight $W_S$ gives a direct control knob for trading accuracy against compression, so the same algorithm can target different parameter-reduction ratios.

Reading between the lines

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

The paper's comparison assumes a 10-epoch baseline, and its baseline accuracies on JAFFE and MUG are far below what deeper, fully converged networks report; a fairer test would compare EvoFER's patches against a converged full-image network, since small inputs converge faster and some of the reported gain may be convergence speed rather than attention.
EvoFER's nose-relative encoding suggests a transfer test the authors did not run: evolve patches on one dataset and apply them without further search to another face dataset; if eyes-and-mouth patches transfer, the method becomes a reusable preprocessing rather than a per-dataset search.
The same patch-selection mechanism could be pointed at other tasks with reliable landmarks, such as identity, gaze, or medical image classification, where a reference anatomy point plays the role of the nose.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: the EA-for-FER-patch-selection idea is new and worth a look, but the paper's headline claim — that EA-selected patches beat the same CNN on full images with 95% fewer parameters — is not yet established. The missing control is not a small omission. Table V compares a full-image baseline against evolved patches only. Since the chromosome search space is initialized randomly and fitness rewards both accuracy and parameter reduction, you cannot tell whether the EA found informative regions or merely that any small non-redundant stack of crops helps this small CNN. For the same reason, the "coincides with human attention" claim is overreach: random or fixed nose-relative crops were not shown for comparison.

What is genuinely new: the nose-relative chromosome encoding, the shift mutation operator, and the stacked-patch representation for FER. The idea is clean, and the visualization of patch drift from hair/background to eyes/mouth in Figures 8–10 is suggestive. Parameter reduction is real and mostly mechanical — smaller inputs shrink the first dense layer — but mechanical is not worthless. The paper is also honest that it does not compare to state-of-the-art and reports standard deviations over 10 runs.

Soft spots, in proportion. First, the baseline is very likely undertrained: 10 epochs for full images, with MUG baseline at 45.8%, far below known results, so the relative gains are inflated. Second, no random-patch control, no fixed-crop control, and no statistical significance tests. Third, the validation split is not described precisely. Fourth, the abstract says the method addresses long training times, while the EA itself takes up to 5 hours; this framing should be fixed. Fifth, no code or data splits are provided.

The fitness function explicitly rewards parameter reduction, so the 95% figure is partly built into the objective. That is not a circularity violation — the test accuracies are out-of-sample — but it undercuts the "emergent discovery" framing. The stress-test note is right: this is a load-bearing gap, not a nitpick.

Who this is for: people working on hybrid EA/CNN systems for resource-constrained recognition, and referees who can push for proper controls. I would bring it to a reading group to discuss experimental standards in evolutionary deep learning. Recommendation: send to peer review with expectation of major revision — the core idea is worth referee time, but the current evidence does not support the central claim as stated.

Referee Report

4 major / 5 minor

Summary. The paper proposes EvoFER, an evolutionary algorithm that encodes, relative to the nose, the size and locations of small patches to extract from a face image; the patches are stacked into a multi-channel input to a fixed CNN. The fitness function (Eq. 1) rewards both high validation accuracy relative to a full-image baseline and a reduction in trainable parameters. Experiments on JAFFE, KDEF, MUG, and RAFD report that the best evolved chromosome improves test accuracy over the same CNN trained on full images by 1.3 to 20.6 percentage points while reducing parameters by 88-98%.

Significance. If the central claim were fully supported, the paper would make a useful practical contribution: a cheap, interpretable preprocessing step that shrinks a CNN's input and parameter count while preserving or improving accuracy, and that identifies facial regions relevant to emotion. The manuscript has real strengths: it uses subject-disjoint train/test splits, repeats each experiment ten times, reports parameter counts and timings, and evaluates on four datasets with varied resolution. However, the empirical design lacks a control for random patch selection, the baseline CNN is likely undertrained, and no inferential statistics are reported. The parameter reduction is also explicitly built into the fitness function, so the 95% figure is partly a statement about the objective rather than an emergent property. These gaps make the strong claims in the abstract and conclusion disproportionate to the evidence.

major comments (4)

[Section IV, Tables V and VI] The central attribution claim is not supported because no random-patch control is included. EvoFER is compared only against a full-image baseline and against its own best evolved chromosome; there is no condition in which the same number of patches of the same size are drawn randomly (or from fixed facial regions) and fed to the same CNN with the same training budget. Without such a control, the observed accuracy gains could be due to the smaller input size or to regularization from cropping, rather than to the evolutionary search over patch locations. The discussion in Section IV that 'the best predictive performance is achieved when pixels around the eyes and mouth are obtained' requires a random or fixed-patch comparison to rule out that any informative crop selection works as well.
[Section III.E and Table V] The full-image baseline is likely undertrained, which undermines the claim of 'no loss in classification accuracy' and the reported improvements. The baseline is trained for only 10 epochs with batch size 8, and the resulting accuracies (45.8% for MUG, 60.0% for JAFFE) are far below published results on these datasets. Because larger input images typically require more epochs to converge, the comparison may favor the patch-based method for reasons unrelated to patch selection. The authors should train the baseline to convergence (or report learning curves showing both models have converged) and should situate both methods relative to published state-of-the-art numbers.
[Tables V and VI] No statistical significance testing is reported, despite the fact that each experiment was repeated ten times. For example, the KDEF comparison (61.9 ± 1.6 vs 63.2 ± 2.9) has heavily overlapping standard deviations, so the claimed 1.3% improvement may not be significant. A paired test across the ten runs, or reporting of per-run differences with confidence intervals, is needed to support the quantitative claims of improvement on all datasets.
[Eq. (1) and Table VI] The parameter reduction is explicitly part of the optimization objective through the term (P_b - P_c)/P_b, and it is also a mechanical consequence of feeding smaller stacked images into a CNN whose first dense layer size depends on the input dimensions. The paper should be careful not to present the 95% reduction as an emergent discovery of the evolutionary process; it is a designed objective. A more informative analysis would report accuracy at matched parameter counts or plot the accuracy-parameter trade-off for random, fixed, and evolved patches.

minor comments (5)

[Figure 2] The caption contains a typo: 'extacted patches' should be 'extracted patches'.
[Section III.D] The word 'dependant' should be 'dependent'.
[Section II.E] The fitness example computes exp(5 × 0.48 + 0.99) ≈ 29.67; it would help to state explicitly that 0.48 is Sc/Sb and 0.99 is (Pb - Pc)/Pb, since the text describes these quantities only just above.
[Abstract and Section IV] The claim that selected patches 'coincide with typical human choices of important features' is supported only by visual inspection of Figures 8-10; a quantitative comparison with human fixation data or with a fixed eye/mouth region baseline would be needed to make this claim.
[Algorithm 1] The random ranges in GenerateX and GenerateY use image width/2 - alpha and image height/2 - beta, which can produce patches whose lower-right corner extends outside the image; the text should clarify how out-of-bounds patches are handled.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the accuracy comparison is out-of-sample, and the parameter-reduction figure is an explicit fitness objective rather than an independent prediction.

full rationale

The central accuracy claim is not circular: EvoFER's fitness (Eq. 1) uses validation accuracy (S_c/S_b), while Table V reports held-out test accuracy, and selection of the best chromosome is based on the validation fitness, not the test set. The parameter-reduction result is indeed encoded in the fitness function as (P_b-P_c)/P_b, and the chromosome bounds (alpha,beta 30-50; 1-4 patches) make large reductions nearly unavoidable, so the 88-98% parameter reductions in Table VI are an optimization objective rather than an emergent discovery. However, the paper transparently presents them as the algorithm's objective, not as a prediction, and the scientific claim that accuracy is retained or improved is tested out-of-sample. No load-bearing self-citation exists: the only self-citation (EDEN, [36]) is a passing remark about an alternative optimizer, and the acknowledgment of the thesis is not an argument. The missing random-patch control and the 10-epoch baseline are experimental confounds or fairness concerns, not circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several user-chosen parameters and domain assumptions. The most important free parameters are the fitness weight W_S and the patch size/count bounds, because they directly shape the reported parameter reduction. The most fragile axiom is that the 10-epoch full-image baseline is a fair reference, since larger images typically need longer training. No new entities are introduced.

free parameters (5)

W_S = 5
Weight of validation accuracy in the fitness function (Eq. 1), chosen by trial runs over {0.1, 5}. It directly controls the tradeoff between accuracy and parameter reduction and is central to what the EA optimizes.
Patch size bounds (alpha, beta) = [30, 50]
Minimum and maximum patch width/height, chosen by hyper-parameter search. These bounds cap the stacked input size and therefore directly determine the achievable parameter reduction; larger bounds would reduce the reported compression.
Patch count bounds = min=1, max=4
User-defined bounds on the number of (x,y) pairs per chromosome. They control the number of stacked patches and thus the input channels and parameter count; the observed parameter reductions depend on these bounds.
EA hyperparameters = population=100, tournament=7, crossover=50%, mutation=50%, generations=15
Selected by a random hyper-parameter search. These affect the quality of the evolved patches but are not directly load-bearing for the reported accuracy; they are standard EA settings.
CNN hyperparameters = epochs=10, batch=8, dropout=0.5, layer sizes
Architecture and training settings chosen from literature and preliminary runs on JAFFE. The choice of 10 epochs is critical because it may undertrain the full-image baseline, influencing the accuracy comparison.

assumptions (5)

domain assumption The baseline CNN trained on full images with 10 epochs is a fair reference for the comparison.
Entered in Section III.E and used in Table V. If the baseline is undertrained, the reported accuracy gains of EvoFER are inflated, which would invalidate the central 'no loss in accuracy' claim as a general statement.
domain assumption Nose detection via OpenCV and DLIB provides reliable reference coordinates across all datasets.
Introduced in Section II.A. The chromosome coordinates are relative to the nose; if nose detection fails or is imprecise, the patch locations are misaligned and the reported performance would degrade.
ad hoc to paper The exponential fitness function form is an appropriate objective for the multi-objective problem.
Equation 1 is chosen without derivation from first principles. It combines accuracy and parameter reduction into a single scalar and any other functional form could change the evolved patch configurations.
domain assumption The subject-exclusive train/test split prevents identity leakage.
Described in Section III.E, following Lopes et al. If subjects appear in both training and testing, the reported accuracy would be optimistically biased.
domain assumption The literature-inspired CNN architecture (3 conv, 2 FC, dropout 0.5) is adequate for FER on these datasets.
Section III.D. The architecture is fixed for both baseline and EvoFER; its suitability affects both, but the baseline is more sensitive to undertraining.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Deep Evolution for Facial Emotion Recognition." pith.science (2026). https://pith.science/paper/2ZORRICS

@misc{pith2026200914194,
  author       = {Pith},
  title        = {Pith review of: Deep Evolution for Facial Emotion Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2ZORRICS}},
  note         = {Machine review of arXiv:2009.14194}
}

read the original abstract

Deep facial expression recognition faces two challenges that both stem from the large number of trainable parameters: long training times and a lack of interpretability. We propose a novel method based on evolutionary algorithms, that deals with both challenges by massively reducing the number of trainable parameters, whilst simultaneously retaining classification performance, and in some cases achieving superior performance. We are robustly able to reduce the number of parameters on average by 95% (e.g. from 2M to 100k parameters) with no loss in classification accuracy. The algorithm learns to choose small patches from the image, relative to the nose, which carry the most important information about emotion, and which coincide with typical human choices of important features. Our work implements a novel form attention and shows that evolutionary algorithms are a valuable addition to machine learning in the deep learning era, both for reducing the number of parameters for facial expression recognition and for providing interpretable features that can help reduce bias.

Figures

Figures reproduced from arXiv: 2009.14194 by the authors.

**Figure 1.** Illustrating the primary idea behind EvoFER. Instead of inputting an entire face image into a CNN to get the prediction, we propose to input patches to achieve similar predictive performance. The extracted patches are stacked together to form a new image. Expression image extracted from [17]. stacked patches (extracted from EvoFER) being inputted into the same CNN. After converting to greyscale, the resolution of th… view at source ↗

**Figure 3.** Illustrating shift mutation on the left and standard mutation on the [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Illustrating the crossover operator swapping the first patch between [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Figures from the paper (5 more)

**Figure 6.** Figure 6: Illustrating the different augmentation techniques which were used to [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 7.** Figure 7: The CNN architecture which is proposed for the experiments. This [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 9.** Figure 9: Illustrating the patches extracted from the best chromosome in [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

**Figure 10.** Figure 10: Illustrating the patches extracted from the best chromosome in [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]

**Figure 8.** Figure 8: Illustrating the patches extracted from the best chromosome in [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 40 canonical work pages

[1]

Facial expression recognition,

Y . Tian, T. Kanade, and J. F. Cohn, “Facial expression recognition,” in Handbook of face recognition . Springer, 2011, pp. 487–519

work page 2011
[2]

The human face as a dynamic tool for social communication,

R. Jack and P. Schyns, “The human face as a dynamic tool for social communication,” Current Biology , vol. 25, no. 14, pp. R621 – R634, 2015

work page 2015
[3]

Face expression recognition: A brief overview of the last decade,

C. D. C?leanu, “Face expression recognition: A brief overview of the last decade,” in 2013 IEEE 8th International Symposium on Applied Computational Intelligence and Informatics , May 2013, pp. 157–161

work page 2013
[4]

Deep learning,

Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, p. 436, 2015

2015
[5]

Handwritten digit recognition with a back- propagation network,

Y . LeCun, B. E. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. E. Hubbard, and L. D. Jackel, “Handwritten digit recognition with a back- propagation network,” in Advances in Neural Information Processing Systems, 1990, pp. 396–404

work page 1990
[6]

Survey on rgb, 3d, thermal, and multimodal approaches for facial expres- sion recognition: History, trends, and affect-related applications,

C. A. Corneanu, M. O. Simon, J. F. Cohn, and S. E. Guerrero, “Survey on rgb, 3d, thermal, and multimodal approaches for facial expres- sion recognition: History, trends, and affect-related applications,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 38, no. 8, pp. 1548–1568, Aug 2016

work page 2016
[7]

Facial Expression Recognition using Convolutional Neural Networks: State of the Art

C. Pramerdorfer and M. Kampel, “Facial expression recognition us- ing convolutional neural networks: State of the art,” arXiv preprint arXiv:1612.02903, 2016

work page Pith review arXiv 2016
[8]

Automatic analysis of facial expressions: the state of the art,

M. Pantic and L. J. M. Rothkrantz, “Automatic analysis of facial expressions: the state of the art,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 22, no. 12, pp. 1424–1445, Dec 2000

work page 2000

Show all 41 references

[9]

Automatic analysis of facial affect: A survey of registration, representation, and recognition,

E. Sariyanidi, H. Gunes, and A. Cavallaro, “Automatic analysis of facial affect: A survey of registration, representation, and recognition,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 37, no. 6, pp. 1113–1133, June 2015

2015
[10]

Survey of the facial expression recognition research,

T. Wu, S. Fu, and G. Yang, “Survey of the facial expression recognition research,” in Advances in Brain Inspired Cognitive Systems , H. Zhang, A. Hussain, D. Liu, and Z. Wang, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2012, pp. 392–402

2012
[11]

Martinez and M

B. Martinez and M. F. Valstar, Advances, Challenges, and Opportunities in Automatic Facial Expression Recognition . Cham: Springer Interna- tional Publishing, 2016, pp. 63–100

2016
[12]

A quick review of deep learning in facial expression,

M. Ghayoumi, “A quick review of deep learning in facial expression,” Journal of Communication and Computer , vol. 14, pp. 34–38, 2017

2017
[13]

Zhang, Facial Expression Recognition Based on Deep Learning: A Survey

T. Zhang, Facial Expression Recognition Based on Deep Learning: A Survey. Cham: Springer International Publishing, 2018, pp. 345–352

2018
[14]

A brief review of facial emotion recognition based on visual information,

B. C. Ko, “A brief review of facial emotion recognition based on visual information,” Sensors, vol. 18, no. 2, p. 401, 2018

2018
[15]

A review on deep learning algorithms for speech and facial emotion recognition,

C. P. Latha and M. Priya, “A review on deep learning algorithms for speech and facial emotion recognition,” APTIKOM Journal on Computer Science and Information Technologies , vol. 1, no. 3, pp. 88–104, 2016

2016
[16]

Genetic algorithms, selection schemes, and the varying effects of noise,

B. L. Miller and D. E. Goldberg, “Genetic algorithms, selection schemes, and the varying effects of noise,” Evolutionary Computation , vol. 4, no. 2, pp. 113–131, Jun. 1996

1996
[17]

The karolinska directed emotional faces ? kdef,

D. Lundqvist, F. A., and A. Ohman, “The karolinska directed emotional faces ? kdef,” 1998

1998
[18]

Coding facial ex- pressions with gabor wavelets,

M. Lyons, S. Akamatsu, M. Kamachi, and J. Gyoba, “Coding facial ex- pressions with gabor wavelets,” inProceedings Third IEEE International Conference on Automatic Face and Gesture Recognition , Apr 1998, pp. 200–205

1998
[19]

The mug facial expression database,

N. Aifanti, C. Papachristou, and A. Delopoulos, “The mug facial expression database,” in 11th International Workshop on Image Analysis for Multimedia Interactive Services , April 2010, pp. 1–4

2010
[20]

Presentation and validation of the radboud faces database,

O. Langner, R. Dotsch, G. Bijlstra, D. H. Wigboldus, S. T. Hawk, and A. Van Knippenberg, “Presentation and validation of the radboud faces database,” Cognition and emotion , vol. 24, no. 8, pp. 1377–1388, 2010

2010
[21]

Image based static facial expression recognition with multiple deep network learning,

Z. Yu and C. Zhang, “Image based static facial expression recognition with multiple deep network learning,” in Proceedings of the 2015 ACM on International Conference on Multimodal Interaction , ser. Interna- tional Conference on Multimodal Interaction ’15. New York, NY , USA:...

2015
[22]

Fusing aligned and non-aligned face information for automatic affect recognition in the wild: A deep learning approach,

B. K. Kim, S. Y . Dong, J. Roh, G. Kim, and S. Y . Lee, “Fusing aligned and non-aligned face information for automatic affect recognition in the wild: A deep learning approach,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition Workshops , June 2016, pp. 1499–1508

2016
[23]

Baseline cnn structure analysis for facial expression recognition,

M. Shin, M. Kim, and D. S. Kwon, “Baseline cnn structure analysis for facial expression recognition,” in 25th IEEE International Symposium on Robot and Human Interactive Communication , Aug 2016, pp. 724–729

2016
[24]

Facial emotion detection using deep learning,

D. Spiers, “Facial emotion detection using deep learning,” 2016

2016
[25]

Head-pose invariant facial expression recognition using con- volutional neural networks,

B. Fasel, “Head-pose invariant facial expression recognition using con- volutional neural networks,” in Proceedings. F ourth IEEE International Conference on Multimodal Interfaces , 2002, pp. 529–534

2002
[26]

Deep learning using linear support vector machines,

Y . Tang, “Deep learning using linear support vector machines,” arXiv preprint arXiv:1306.0239, 2013

2013 arXiv
[27]

Facial expression recognition using deep convolutional neural networks,

D. V . Sang, N. V . Dat, and D. P. Thuan, “Facial expression recognition using deep convolutional neural networks,” International Conference on Knowledge and Systems Engineering , 2017

2017
[28]

Facial expression recog- nition based on transfer learning from deep convolutional networks,

M. Xu, W. Cheng, Q. Zhao, L. Ma, and F. Xu, “Facial expression recog- nition based on transfer learning from deep convolutional networks,” in 11th International Conference on Natural Computation , Aug 2015, pp. 702–708

2015
[29]

Devel- opment of deep learning-based facial expression recognition system,

H. Jung, S. Lee, S. Park, B. Kim, J. Kim, I. Lee, and C. Ahn, “Devel- opment of deep learning-based facial expression recognition system,” in 21st Korea-Japan Joint Workshop on Frontiers of Computer Vision , Jan 2015, pp. 1–4

2015
[30]

Cross-database facial expression recognition based on ﬁne-tuned deep convolutional network,

M. V . Zavarez, R. F. Berriel, and T. Oliveira-Santos, “Cross-database facial expression recognition based on ﬁne-tuned deep convolutional network,” in 30th SIBGRAPI Conference on Graphics, Patterns and Images, Oct 2017, pp. 405–412

2017
[31]

Automatic facial expression recognition,

H. G. Valero, “Automatic facial expression recognition,” 2016

2016
[32]

A deep neural network-driven feature learning method for multi-view facial expression recognition,

T. Zhang, W. Zheng, Z. Cui, Y . Zong, J. Yan, and K. Yan, “A deep neural network-driven feature learning method for multi-view facial expression recognition,” IEEE Transactions on Multimedia , vol. 18, no. 12, pp. 2528–2536, Dec 2016

2016
[33]

Convolutional neural networks for facial expression recognition,

S. Alizadeh and A. Fazel, “Convolutional neural networks for facial expression recognition,” arXiv preprint arXiv:1704.06756 , 2017

2017 arXiv
[34]

Training deep networks for facial expression recognition with crowd-sourced label distribution,

E. Barsoum, C. Zhang, C. C. Ferrer, and Z. Zhang, “Training deep networks for facial expression recognition with crowd-sourced label distribution,” in Proceedings of the 18th ACM International Conference on Multimodal Interaction , ser. International Conference on Multimodal I...

2016
[35]

Facial expression recognition with convolutional neural networks,

A. Raghuvanshi and V . Choksi, “Facial expression recognition with convolutional neural networks,” 2016

2016
[36]

Eden: Evolutionary deep networks for efﬁcient machine learning,

E. Dufourq and B. A. Bassett, “Eden: Evolutionary deep networks for efﬁcient machine learning,” in 2017 Pattern Recognition Association of South Africa and Robotics and Mechatronics , Nov 2017, pp. 110–115

2017
[37]

S. Zhou, Y . Liang, J. Wan, and S. Z. Li, Facial Expression Recognition Based on Multi-scale CNNs . Cham: Springer International Publishing, 2016, pp. 503–510

2016
[38]

Facial expression recognition using a hybrid cnn-sift aggregator,

M. Al-Shabi, W. P. Cheah, and T. Connie, “Facial expression recognition using a hybrid cnn-sift aggregator,” CoRR, vol. abs/1608.02833, 2016

2016 arXiv
[39]

Group-level emotion recognition using transfer learning from face identiﬁcation,

A. Rassadin, A. Gruzdev, and A. Savchenko, “Group-level emotion recognition using transfer learning from face identiﬁcation,” in Pro- ceedings of the 19th ACM International Conference on Multimodal Interaction, ser. International Conference on Multimodal Interaction
[40]

Facial expression recognition with convolutional neural networks: Coping with few data and the training sample order,

A. T. Lopes, E. de Aguiar, A. F. D. Souza, and T. Oliveira-Santos, “Facial expression recognition with convolutional neural networks: Coping with few data and the training sample order,” Pattern Recognition, vol. 61, pp. 610 – 628, 2017

2017
[2017]

New York, NY , USA: ACM, 2017, pp. 544–548

2017

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.

[1] [1]

Facial expression recognition,

Y . Tian, T. Kanade, and J. F. Cohn, “Facial expression recognition,” in Handbook of face recognition . Springer, 2011, pp. 487–519

work page 2011

[2] [2]

The human face as a dynamic tool for social communication,

R. Jack and P. Schyns, “The human face as a dynamic tool for social communication,” Current Biology , vol. 25, no. 14, pp. R621 – R634, 2015

work page 2015

[3] [3]

Face expression recognition: A brief overview of the last decade,

C. D. C?leanu, “Face expression recognition: A brief overview of the last decade,” in 2013 IEEE 8th International Symposium on Applied Computational Intelligence and Informatics , May 2013, pp. 157–161

work page 2013

[4] [4]

Deep learning,

Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, p. 436, 2015

2015

[5] [5]

Handwritten digit recognition with a back- propagation network,

Y . LeCun, B. E. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. E. Hubbard, and L. D. Jackel, “Handwritten digit recognition with a back- propagation network,” in Advances in Neural Information Processing Systems, 1990, pp. 396–404

work page 1990

[6] [6]

Survey on rgb, 3d, thermal, and multimodal approaches for facial expres- sion recognition: History, trends, and affect-related applications,

C. A. Corneanu, M. O. Simon, J. F. Cohn, and S. E. Guerrero, “Survey on rgb, 3d, thermal, and multimodal approaches for facial expres- sion recognition: History, trends, and affect-related applications,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 38, no. 8, pp. 1548–1568, Aug 2016

work page 2016

[7] [7]

Facial Expression Recognition using Convolutional Neural Networks: State of the Art

C. Pramerdorfer and M. Kampel, “Facial expression recognition us- ing convolutional neural networks: State of the art,” arXiv preprint arXiv:1612.02903, 2016

work page Pith review arXiv 2016

[8] [8]

Automatic analysis of facial expressions: the state of the art,

M. Pantic and L. J. M. Rothkrantz, “Automatic analysis of facial expressions: the state of the art,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 22, no. 12, pp. 1424–1445, Dec 2000

work page 2000