pith. sign in

arxiv: 2510.21833 · v1 · submitted 2025-10-22 · 💻 cs.CV

Towards Accurate and Efficient Waste Image Classification: A Hybrid Deep Learning and Machine Learning Approach

Ngoc-Bao-Quang Nguyen , Tuan-Minh Do , Cong-Tam Phan , Thi-Thu-Hong Phan This is my paper

Pith reviewed 2026-05-18 05:25 UTC · model grok-4.3

classification 💻 cs.CV
keywords waste image classificationhybrid deep learningmachine learning classifiersfeature extractionTrashNetGarbage Classificationimage-based recyclingfeature selection
0
0 comments X

The pith

A hybrid of deep feature extraction and classical classifiers reaches 100% accuracy on waste image datasets while shrinking features by over 95%.

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

This paper compares three strategies for sorting garbage from photos: handcrafted machine learning, full deep learning networks such as ResNet variants and EfficientNetV2S, and a hybrid that uses deep networks only to pull out image features before handing them to simpler classifiers like support vector machines or logistic regression. Across TrashNet, Garbage Classification, and a corrected Household Garbage set the hybrid records the highest scores, reaching perfect accuracy on two collections and 99.87 percent on the third. The work further shows that trimming the extracted features by more than 95 percent leaves accuracy unchanged yet speeds training and inference. These outcomes matter because reliable, lightweight waste recognition can support automated recycling systems even where computing resources are scarce.

Core claim

The study establishes that a hybrid strategy, in which deep learning models serve only as feature extractors and classical machine learning classifiers perform the final decision, consistently delivers higher accuracy than either pure machine learning or pure deep learning on public waste image datasets. On the TrashNet and a corrected Household Garbage set the hybrid reaches 100 percent accuracy, while on the Garbage Classification set it attains 99.87 percent, exceeding published state-of-the-art figures. In addition, reducing the feature space by more than 95 percent through selection leaves accuracy intact and lowers both training and inference time.

What carries the argument

The hybrid pipeline that pairs deep convolutional networks for feature extraction with classical classifiers such as support vector machines and logistic regression.

If this is right

  • The hybrid method outperforms both standalone machine learning and deep learning models on all tested waste datasets.
  • Feature selection cuts dimensionality by over 95 percent while preserving accuracy.
  • Training and inference become faster, supporting deployment on devices with limited resources.
  • More reliable benchmarks are provided for future waste classification research.

Where Pith is reading between the lines

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

  • Similar hybrids could improve accuracy in other image classification domains where pure deep models struggle with small or noisy datasets.
  • The reported gains may partly stem from the dataset cleaning step, suggesting that label quality deserves equal attention to model architecture in future work.
  • On-device sorting applications become more practical once feature count is reduced this sharply.

Load-bearing premise

The manual correction of 43 mislabels in the Household Garbage Dataset produces a cleaner but still representative test set.

What would settle it

Evaluating the same hybrid model on an independent waste image collection that has not undergone manual label correction and checking whether accuracy remains above 95 percent.

Figures

Figures reproduced from arXiv: 2510.21833 by Cong-Tam Phan, Ngoc-Bao-Quang Nguyen, Thi-Thu-Hong Phan, Tuan-Minh Do.

Figure 1
Figure 1. Figure 1: Overview of proposed approach for garbage image classification [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Sample images from the Garbage Classification dataset [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sample images from the Household Garbage dataset [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sample images from the TrashNet dataset 4.2. Results and discussion This section presents a comprehensive evaluation of the proposed classification frameworks across the three selected datasets. To ensure a fair and consistent comparison, a uniform preprocessing pipeline was applied to all images prior to model training and evaluation. Specifically, each image was resized to a resolution of 400×400 pixels … view at source ↗
Figure 5
Figure 5. Figure 5: Summary of the best performance from each approach across all datasets. [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: An example of a mislabeled image found during the inspection. [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of our proposed hybrid model with previous methods (SOTA) across three datasets: [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
read the original abstract

Automated image-based garbage classification is a critical component of global waste management; however, systematic benchmarks that integrate Machine Learning (ML), Deep Learning (DL), and efficient hybrid solutions remain underdeveloped. This study provides a comprehensive comparison of three paradigms: (1) machine learning algorithms using handcrafted features, (2) deep learning architectures, including ResNet variants and EfficientNetV2S, and (3) a hybrid approach that utilizes deep models for feature extraction combined with classical classifiers such as Support Vector Machine and Logistic Regression to identify the most effective strategy. Experiments on three public datasets - TrashNet, Garbage Classification, and a refined Household Garbage Dataset (with 43 corrected mislabels)- demonstrate that the hybrid method consistently outperforms the others, achieving up to 100% accuracy on TrashNet and the refined Household set, and 99.87% on Garbage Classification, thereby surpassing state-of-the-art benchmarks. Furthermore, feature selection reduces feature dimensionality by over 95% without compromising accuracy, resulting in faster training and inference. This work establishes more reliable benchmarks for waste classification and introduces an efficient hybrid framework that achieves high accuracy while reducing inference cost, making it suitable for scalable deployment in resource-constrained environments.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 2 minor

Summary. The manuscript compares three paradigms for automated waste image classification: (1) classical ML with handcrafted features, (2) deep learning models including ResNet variants and EfficientNetV2S, and (3) a hybrid approach that extracts features from deep models and feeds them to classical classifiers such as SVM and Logistic Regression. Experiments are reported on TrashNet, the Garbage Classification dataset, and a refined Household Garbage Dataset in which 43 mislabels were manually corrected. The central empirical claim is that the hybrid method consistently outperforms the other paradigms, reaching up to 100% accuracy on TrashNet and the refined Household set and 99.87% on Garbage Classification, thereby surpassing prior state-of-the-art results; an additional result is that feature selection reduces dimensionality by more than 95% while preserving accuracy and improving training/inference speed.

Significance. If the performance claims prove robust under proper statistical controls and unbiased dataset curation, the work would supply useful empirical benchmarks for waste classification and demonstrate that hybrid DL-feature + classical-classifier pipelines can deliver high accuracy at lower inference cost than end-to-end deep models. The dimensionality-reduction result is a concrete practical contribution for resource-constrained deployment. The study is strengthened by its use of three public datasets and its explicit comparison across ML, DL, and hybrid regimes.

major comments (3)
  1. [Dataset description and refinement] Dataset preparation (refined Household Garbage Dataset): the manual correction of 43 mislabels is stated without any protocol details (blinding, number of annotators, inter-rater agreement, or whether corrections were informed by preliminary model outputs). Because the headline 100% accuracy on this refined set is load-bearing for the claim that the hybrid method surpasses SOTA, the absence of these safeguards leaves open the possibility that the test distribution was inadvertently made easier or biased toward the proposed method.
  2. [Experimental setup and results] Experimental protocol (§4 / results tables): no train-test split ratios, random seeds, cross-validation scheme, or number of independent runs are reported, and all accuracy figures lack error bars or standard deviations. These omissions make it impossible to judge whether the reported gains over baselines are statistically reliable or merely the result of a single favorable split.
  3. [Comparison with prior work] SOTA comparison: the assertion that the hybrid method surpasses state-of-the-art benchmarks is not supported by direct re-implementation or identical-split replication of the exact prior methods cited. Without such controls, the performance advantage cannot be isolated from differences in data partitioning or hyper-parameter tuning.
minor comments (2)
  1. [Abstract] The abstract should explicitly name the precise hybrid configuration (feature extractor + classifier) that produced each peak accuracy figure.
  2. [Results tables] Tables reporting accuracy should include the dimensionality after feature selection and the corresponding training/inference times to make the efficiency claim quantitative.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We have carefully addressed each major comment below and revised the manuscript to enhance reproducibility, transparency, and the robustness of our empirical claims. We believe these changes strengthen the paper without altering its core contributions.

read point-by-point responses

  1. Referee: Dataset preparation (refined Household Garbage Dataset): the manual correction of 43 mislabels is stated without any protocol details (blinding, number of annotators, inter-rater agreement, or whether corrections were informed by preliminary model outputs). Because the headline 100% accuracy on this refined set is load-bearing for the claim that the hybrid method surpasses SOTA, the absence of these safeguards leaves open the possibility that the test distribution was inadvertently made easier or biased toward the proposed method.

    Authors: We appreciate this important point regarding transparency and potential bias. The 43 corrections were made by two authors through independent visual inspection of the images, relying solely on obvious labeling errors (e.g., clear mismatches between image content and label) without any reference to model outputs or predictions. In the revised manuscript, we have added a dedicated paragraph in Section 3.2 describing the full protocol: the process was blinded (annotators had no access to preliminary results), inter-rater agreement reached Cohen's kappa of 0.87, and any disagreements were resolved by joint review. We have also released both the original and corrected label files to enable independent verification. These additions directly address the concern while preserving the integrity of the refined dataset. revision: yes

  2. Referee: Experimental protocol (§4 / results tables): no train-test split ratios, random seeds, cross-validation scheme, or number of independent runs are reported, and all accuracy figures lack error bars or standard deviations. These omissions make it impossible to judge whether the reported gains over baselines are statistically reliable or merely the result of a single favorable split.

    Authors: We agree that these details are essential for evaluating statistical reliability. In the revised Section 4, we now explicitly state that all experiments used an 80/20 train-test split with a fixed random seed of 42 for the primary results. Hyperparameter tuning was performed via 5-fold cross-validation on the training portion. To quantify variability, we conducted 5 independent runs with different seeds (42, 123, 456, 789, 1011) and report mean accuracy ± standard deviation in all tables. Error bars have been added to the corresponding figures. These revisions allow readers to assess whether the observed improvements are consistent across runs. revision: yes

  3. Referee: SOTA comparison: the assertion that the hybrid method surpasses state-of-the-art benchmarks is not supported by direct re-implementation or identical-split replication of the exact prior methods cited. Without such controls, the performance advantage cannot be isolated from differences in data partitioning or hyper-parameter tuning.

    Authors: We acknowledge that direct re-implementation with identical splits provides the strongest evidence. Our original comparisons used the accuracy figures as reported in the cited works on the same public datasets. In the revision, we have added partial re-implementations for two key prior methods (where open-source code was available), ensuring the same data splits and reporting protocol as our experiments. For the remaining works without public code, we have clarified in the text the potential confounding factors and noted that our hybrid results remain competitive even under these constraints. We maintain that the hybrid paradigm demonstrates clear practical advantages, but we have tempered the SOTA claim language to reflect these limitations. revision: partial

Circularity Check

0 steps flagged

No derivation chain present; empirical results on public datasets are self-contained

full rationale

The paper reports direct experimental comparisons of handcrafted ML, DL architectures, and hybrid feature-extraction-plus-classifier pipelines on three public datasets (TrashNet, Garbage Classification, refined Household Garbage). No equations, first-principles derivations, or predictive models that reduce to fitted parameters are claimed. All performance numbers (including the 100% accuracy after correcting 43 labels) are measured outcomes on held-out test splits rather than outputs of any internal derivation that could be circular by construction. The label-correction step is a preprocessing choice whose fairness can be questioned on other grounds, but it does not create a self-referential loop in any claimed result.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central empirical claims rest on the assumption that the three public datasets are representative of real-world waste images and that the 43 manual label corrections do not introduce selection bias. No new mathematical axioms or invented physical entities are introduced.

axioms (1)
  • domain assumption The refined Household Garbage Dataset after correcting 43 mislabels remains a fair test of generalization to unseen waste images.
    Stated in the abstract when reporting 100% accuracy on the refined set.

pith-pipeline@v0.9.0 · 5759 in / 1346 out tokens · 31254 ms · 2026-05-18T05:25:47.270675+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

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

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

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages · 1 internal anchor

  1. [1]

    J. Fu, Analysis of machine learning recognition method for garbage classification, The- oretical and Natural Science 120 (2025) 139–145, available at:https://doi.org/10.5 4254/2753-8818/2025.AD26349

    work page 2025

  2. [2]

    Aghilan, M

    M. Aghilan, M. Arun Kumar, T. S. Mohammed Aafrid, A. Nirmal Kumar, S. Muthulak- shmi, Garbage waste classification using supervised deep learning techniques, IJETIE 6 (3), available at:https://ssrn.com/abstract=3563564(2020)

    work page 2020

  3. [3]

    Huang, J

    G. Huang, J. He, Z. Xu, G. Huang, A combination model based on transfer learning for waste classification, Concurrency and Computation: Practice and Experience 32 (19) (2020) e5751, available at:https://doi.org/10.1002/cpe.5751

    work page doi:10.1002/cpe.5751 2020

  4. [4]

    Wulansari, A

    A. Wulansari, A. Setyanto, E. T. Luthfi, A convolutional neural network based on trans- fer learning for medical waste classification during pandemic covid-19, in: Proc. 2022 International Seminar on Intelligent Technology and Its Applications (ISITIA), IEEE, 2022, pp. 1–6, available at:https://doi.org/10.1109/ISITIA56226.2022.9855374

    work page doi:10.1109/isitia56226.2022.9855374 2022

  5. [5]

    J. Wang, Application research of image classification algorithm based on deep learning in household garbage sorting, Heliyon 10 (9) (2024) e29966, available at:https://do i.org/10.1016/j.heliyon.2024.e29966

    work page doi:10.1016/j.heliyon.2024.e29966 2024

  6. [6]

    G. Celik, Multi-layer feature fusion for high-accuracy solid waste classification using a hybrid deep learning model, The Visual ComputerAvailable at:https://link.sprin ger.com/article/10.1007/s00371-025-04031-3(2025)

    work page doi:10.1007/s00371-025-04031-3(2025 2025

  7. [7]

    Nahiduzzaman, M

    M. Nahiduzzaman, M. Rahman, M. Uddin, M. Islam, An automated waste classification system using deep learning techniques, Knowledge-Based Systems 310 (2025) 113028, available at:https://www.sciencedirect.com/science/article/pii/S095070512 5000760

    work page 2025

  8. [8]

    L. Li, R. Wang, M. Zou, F. Guo, Y. Ren, Enhanced resnet-50 for garbage classification: Feature fusion and depth-separable convolutions, PLOS ONE 20 (1) (2025) e0317999, available at:https://doi.org/10.1371/journal.pone.0317999

    work page doi:10.1371/journal.pone.0317999 2025

  9. [9]

    S. V. T. Dao, T. H. Tran, T. H. Pham, T. Q. Nguyen, Integrating artificial intelligence for sustainable waste management, Cleaner Waste Systems 15 (2025) 100345, available at:https://www.sciencedirect.com/science/article/pii/S2589471425000270

    work page 2025

  10. [10]

    A. P. Puspaningrum, S. N. Endah, P. S. Sasongko, R. Kusumaningrum, Khadijah, Ris- miyati, Waste classification using support vector machine with sift–pca feature extrac- tion, in: Proc. 2020 4th International Conference on Informatics and Computational Sciences (ICICoS), IEEE, 2020, pp. 56–61, available at:https://doi.org/10.1109/IC ICoS50035.2020.9298982

    work page doi:10.1109/ic 2020

  11. [11]

    C. Shi, C. Tan, T. Wang, L. Wang, A waste classification method based on a multilayer hybrid convolution neural network, Applied Sciences 11 (18) (2021) 8572, available at: https://doi.org/10.3390/app11188572. 28

    work page doi:10.3390/app11188572 2021

  12. [12]

    Lilhore, S

    U. Lilhore, S. Simaiya, S. Dalal, R. Damasevicius, A smart waste classification model using hybrid cnn-lstm with transfer learning, Multimedia Tools and Applications 83 (2023) 29505–29529, available at:https://doi.org/10.1007/s11042-023-16677-z

    work page doi:10.1007/s11042-023-16677-z 2023

  13. [13]

    Mehedi, M

    M. Mehedi, M. Rahman, M. Kabir, S. Hossain, A transfer learning approach for efficient classification of waste materials, in: 2023 IEEE Computing Conference on Wireless Communication and Computing (CCWC), 2023, available at:https://doi.org/10.1 109/CCWC57344.2023.10099127

    work page arXiv 2023

  14. [14]

    Sharma, A

    A. Sharma, A. Keshri, A. Kumar, R. K. Yadav, Garbage classification with deep learning techniques, in: 2023 International Conference on Computational Intelligence and Sustainable Engineering Solutions (CISES), IEEE, 2023, available at:https: //ieeexplore.ieee.org/document/10183566

    work page arXiv 2023

  15. [15]

    Kunwar, Managing household waste through transfer learning, Industrial and Domes- tic Waste Management 4 (1) (2024) 14–22, available at:https://doi.org/10.53623 /idwm.v4i1.408

    S. Kunwar, Managing household waste through transfer learning, Industrial and Domes- tic Waste Management 4 (1) (2024) 14–22, available at:https://doi.org/10.53623 /idwm.v4i1.408

    work page 2024

  16. [16]

    A. S. M. Sayem, M. A. Tareq, M. I. Afjal, S. M. Tuli, M. M. Rahman, Enhancing waste sorting and recycling efficiency: robust deep learning-based approach for classification and detection, Neural Computing and ApplicationsAvailable at:https://link.sprin ger.com/article/10.1007/s00521-024-10855-2(2024)

    work page doi:10.1007/s00521-024-10855-2(2024 2024

  17. [17]

    Rother, V

    C. Rother, V. Kolmogorov, A. Blake, Grabcut: Interactive foreground extraction using iterated graph cuts, ACM Transactions on Graphics 23 (3) (2004) 309–314, available at: https://dl.acm.org/doi/10.1145/1186562.1015720

    work page doi:10.1145/1186562.1015720 2004

  18. [18]

    T.-T.-H. Phan, T. T. T. Hai, V. T. H. L., V. H., T. T. N., Comparative study on vision based rice seed varieties identification, in: Proc. 7th Int. Conf. Knowledge and Systems Engineering (KSE), 2015, pp. 377–382, available at:https://ieeexplore.ieee.org/ document/7371816

    work page arXiv 2015

  19. [19]

    Liu, J.-M

    P. Liu, J.-M. Guo, K. Chamnongthai, H. Prasetyo, Fusion of color histogram and lbp- based features for texture image retrieval and classification, Information Sciences 390 (2017) 95–111, available at:https://doi.org/10.1016/j.ins.2017.01.025

    work page doi:10.1016/j.ins.2017.01.025 2017

  20. [20]

    X. Y. Gong, H. Su, D. Xu, T. Li, X. Zhu, S. Wang, X. Zuo, J. Chen, An overview of contour detection approaches, International Journal of Automation and Computing 15 (2018) 656–672, available at:https://link.springer.com/article/10.1007/s11633 -018-1117-z

    work page doi:10.1007/s11633 2018

  21. [21]

    Hu, Visual pattern recognition by moment invariants, IRE Transactions on Infor- mation Theory 8 (2) (1962) 179–187, available at:https://ieeexplore.ieee.org/do cument/1057692

    M.-K. Hu, Visual pattern recognition by moment invariants, IRE Transactions on Infor- mation Theory 8 (2) (1962) 179–187, available at:https://ieeexplore.ieee.org/do cument/1057692

    work page arXiv 1962

  22. [22]

    R. M. Haralick, K. Shanmugam, I. Dinstein, Textural features for image classification, IEEE Transactions on Systems, Man, and Cybernetics 3 (6) (1973) 610–621, available at:https://ieeexplore.ieee.org/document/4309314. 29

    work page arXiv 1973

  23. [23]

    Ojala, M

    T. Ojala, M. Pietikäinen, D. Harwood, Performance evaluation of texture measures with classification based on kullback discrimination of distributions, in: Proc. 12th IAPR Int. Conf. Pattern Recognition (ICPR), Vol. 1, 1994, pp. 582–585, available at:https: //ieeexplore.ieee.org/document/576366

    work page 1994

  24. [24]

    Rublee, V

    E. Rublee, V. Rabaud, K. Konolige, G. Bradski, Orb: An efficient alternative to sift or surf, in: Proc. IEEE Int. Conf. Computer Vision, 2011, pp. 2564–2571, available at: https://ieeexplore.ieee.org/document/6126544

    work page arXiv 2011

  25. [25]

    D. G. Lowe, Object recognition from local scale-invariant features, in: Proc. 7th IEEE Int. Conf. Computer Vision, Vol. 2, 1999, pp. 1150–1157, available at:https://ieeexp lore.ieee.org/document/790410/authors#authors

    work page 1999

  26. [26]

    Oliva, A

    A. Oliva, A. Torralba, Modeling the shape of the scene: A holistic representation of the spatial envelope, International Journal of Computer Vision 42 (2001) 145–175, available at:https://link.springer.com/article/10.1023/A:1011139631724

    work page doi:10.1023/a:1011139631724 2001

  27. [27]

    D. W. Hosmer, S. Lemeshow, R. X. Sturdivant, Applied Logistic Regression, 3rd Edition, Wiley, Hoboken, NJ, USA, 2013, available at:https://onlinelibrary.wiley.com/do i/book/10.1002/9781118548387

    work page doi:10.1002/9781118548387 2013

  28. [28]

    Cover, P

    T. Cover, P. Hart, Nearest neighbor pattern classification, IEEE Transactions on Infor- mation Theory 13 (1) (1967) 21–27, available at:https://ieeexplore.ieee.org/do cument/1053964

    work page arXiv 1967

  29. [29]

    Support-vector networks,

    C. Cortes, V. Vapnik, Support-vector networks, Machine Learning 20 (3) (1995) 273– 297, available at:https://doi.org/10.1007/BF00994018

    work page doi:10.1007/bf00994018 1995

  30. [30]

    Charbuty, A

    B. Charbuty, A. Abdulazeez, Classification based on decision tree algorithm for machine learning, Journal of Applied Science and Technology Trends (2021) 20–28Available at: https://www.researchgate.net/publication/350386944_Classification_Based_ on_Decision_Tree_Algorithm_for_Machine_Learning

    work page arXiv 2021

  31. [31]

    Machine Learning , author =

    L. Breiman, Random forests, Machine Learning 45 (1) (2001) 5–32, available at:https: //doi.org/10.1023/A:1010933404324

    work page doi:10.1023/a:1010933404324 2001

  32. [32]

    T. Chen, C. Guestrin, Xgboost: A scalable tree boosting system, in: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016, pp. 785–794, available at:https://dl.acm.org/doi/10.1145/2939672 .2939785

    work page doi:10.1145/2939672 2016

  33. [33]

    G. Ke, Q. Meng, T. Finley, W. Wang, W. Chen, W. Ma, Q. Ye, T.-Y. Liu, Lightgbm: A highly efficient gradient boosting decision tree, in: Advances in Neural Information Processing Systems 30 (NIPS 2017), 2017, available at:https://papers.nips.cc/pap er/2017/hash/6449f44a102fde848669bdd9eb6b76fa-Abstract.html

    work page 2017

  34. [34]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 2016, pp. 770–778, available at:https://ieeexplore.ieee.org/document/7780459. 30

    work page arXiv 2016

  35. [35]

    T.-Y. Lin, P. Goyal, R. Girshick, K. He, P. Dollár, Focal loss for dense object detection, in: Proc. IEEE Int. Conf. Computer Vision (ICCV), 2017, pp. 2980–2988, available at: https://doi.org/10.1109/ICCV.2017.324

    work page doi:10.1109/iccv.2017.324 2017

  36. [36]

    M. Tan, Q. V. Le, Efficientnet: Rethinking model scaling for convolutional neural net- works, in: Proc. Int. Conf. Machine Learning (ICML), 2019, pp. 6105–6114, available at:https://doi.org/10.48550/arXiv.1905.11946

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1905.11946 2019

  37. [37]

    Mohamed, Garbage classification dataset, Kaggle, [Online]

    M. Mohamed, Garbage classification dataset, Kaggle, [Online]. Available:https://ww w.kaggle.com/datasets/mostafaabla/garbage-classification(2020)

    work page 2020

  38. [38]

    S.Kunwar, Householdgarbageclassificationdataset, Kaggle, [Online].Available:https: //www.kaggle.com/datasets/sumn2u/garbage-classification-v2(2021)

    work page 2021

  39. [39]

    Ozkefe, Trashnet, Kaggle, [Online]

    F. Ozkefe, Trashnet, Kaggle, [Online]. Available:https://www.kaggle.com/dataset s/feyzazkefe/trashnet(2021). 31

    work page 2021