A Comprehensive Review of Automated Waste Classification and Waste-to-Resource Systems

A Comprehensive Review of Automated Waste Classification and Waste-to-Resource Systems

Abhaynath T G, Abhijith R S, Nidhi AnilKumar, P V Sauravi, Dr. Manju S Nair, Dr. Sabeena K, Chinchu M Pillai

College of Engineering Chengannur, Kerala, India

AbstractMachine learning based waste classication,the ma- chine learning recognition of waste streams from images and sensor information is a keystone technology for contemporary resource recovery systems. In the past ve years, research has taken off: data sets have multiplied, convolutional and attention-based neural networks have been used in detection and classication applications, and system-level platforms have been architected to transform streams of waste into marketable resources. This report integrates results from eight uploaded papers (ranging from dataset construction, CNN- and YOLO- based detectors, systematic reviews of the literature, to industrial- symbiosis platforms) into one uniform survey. Methodological decisions (deep learning, feature-based approaches, one-stage detectors), properties and restrictions of datasets, performance metrics, indicated ranges of reported performance, application situations (MRFs, smart bins, robotic pickers), and social aspects (health, privacy, economics) are discussed. We also offer compar- ative tables, actionable advice for practitioners and researchers, and a ranked roadmap of short-, mid- and long-term directions for research. Our intention is to deliver a single document summarizing the uploaded work and placing it in a dened, actionable context for subsequent research and deployment.

Index TermsWaste classication, deep learning, YOLO, datasets, industrial symbiosis, smart bins, computer vision, trans- fer learning, domain adaptation.

1. Introduction

   THE Automated sorting and classication of waste are essential facilitators of new waste management infras-

   tructure and the circular economy. As consumption and urban- ization rise, manual sorting capacity becomes a limiting factor and a source of human health hazard. Automated vision-based systems have the potential to provide greater throughput, re- duced error rates, and improved safety than manual processes. In addition, coupling classication with resource matching platforms has the potential to allow industrial consumers and waste producers to identify economically sound reuse chan- nels for secondary raw materials. The eight uploaded studies employed within this survey range from level one datasets and algorithms all the way to system level matchmaking platforms and systematic reviews of the discipline, and therefore form a solid foundation upon which to form a consolidated survey. This preface gives background on the economic and ecological rationale for automated waste sorting, discusses the scope of this article, and presents the rest of the structure. Global waste production keeps increasing: recent projections forecast mu- nicipal solid waste growth to several billion tonnes per year in

   future decades, making automation inevitable. Automated sort- ing enables several phases of the value chain: front-end source segregation (intelligent bins), middle-stage material recovery (MRFs and robotic sorters), and back-end matchmaking (in- dustrial symbiosis platforms). Scope and aims: this paper integrates the eight studies uploaded dataset publications, al- gorithmic experiments with CNNs and YOLO-style detectors, and system/platform overviews like industrial symbiosis web tools and combines them with pertinent supporting literature to yield (i) an organized comparison of methods and datasets,

   (ii) working advice for classier implementation and testing,

   (iii) an overview of practical application, and (iv) a prioritized research agenda. Organization: Section II discusses the eight studies comprehensively and categorizes them thematically. Section III consolidates algorithmic families and preprocessing decisions. Section IV consolidates dataset properties and con- straints. Section V presents comparative performance analysis and summarizes evaluation protocols. Section VI enumerates application scenarios and deployment issues. Section VII ex- plores limitations and social implications. Section IX presents a future oriented research agenda. The paper concludes in Section X.

2. Literature Review

   Current research on automatic waste sorting focuses on the use of machine learning and deep learning to enhance recy- cling effectiveness. Convolutional Neural Networks (CNNs) and YOLO-based object detectors have persistently surpassed conventional approaches, allowing for correct, real-time de- tection of recyclable materials. Scholars also pointed to the signicance of stable datasets, stating that class imbalance and lack of diversity are still the main challenges. In addition to classication, system-level strategies like industrial symbiosis platforms facilitate the transformation of waste streams into meaningful resources. Generally, the literature indicates that combining advanced AI models with sustainable business frameworks can greatly improve contemporary waste manage- ment systems. The papers can be classied into three broad categories: (A) systematic review and dataset analysis, (B) algorithmic/deployment experiments (CNN and YOLO detec- tors), and (C) system/platform design for industrial symbiosis and resource matching. Following are in-depth summaries that highlight methods, datasets, experimental results, and practical lessons per paper.

   1. [P1] Systematic Review of AI Based Automated Waste Classication (Sensors 2025)

      This article is a systematic literature review (SLR) of AI- based waste classication methods from 20202025. Almost 100 papers were reviewed, and over a dozen public datasets were itemized. The key contributions are:

      • Taxonomy of methods: Categorization of methods into traditional machine learning, deep learning (CNNs), fu- sion systems, and sensor fusion. Convolutional neural networks and transfer learning are predominant in recent research.

      • Dataset critique: Recognition of limitations of datasets like imbalance, small sample size, missing multi-object scenes, and non-uniform class taxonomies.

      • Deployment gap: Emphasizing the gap between high reported accuracies on handpicked datasets and perfor- mance in actual industrial environments.

      • Roadmap: A study agenda focusing on standardized benchmarks, multimodal fusion, and privacy-preserving collaborative training.

        This SLR acts as a structural foundation for follow-up studies by determining common trends and gaps.

   2. [P2] YOLOv8-Based Real-Time Household Waste Detec- tion (Sustainability 2025)

      This paper presents a YOLOv8-based pipeline for household waste detection. Key distinguishing elements are:

      • Dataset: A 17-class household waste dataset (around 3,700 images) with occlusion and multi-object compo- sitions for simulating real-world conditions.

      • Model innovations: YOLOv8 backbone enhanced with attention modules (CBAM and SE) and an extensive augmentation pipeline (Mosaic, MixUp, random erasing).

      • Results: Reached an mAP of

        sim89.5

      • Practical notes: Latencyaccuracy trade-offs, quantization-aware training, and bounding-box post- processing discussion.

        The research shows that well-engineered one-stage etectors are feasible for edge deployment.

   3. [P3] Trash Detection using CNNs (Conference 2024)

      This work introduces a CNN-based classication pipeline with the following contributions:

      • Pipeline: Transfer learning with MobileNet/ResNet back- bones, class rebalancing, and test-time augmentation.

      • Results: Obtained

        sim89.9

      • Dataset curation: Focus on uniform labeling and class- specic augmentation. ¡/itemize¿.

      • Ablations: Showed the efcacy of pretraining on general datasets prior to domain-specic ne-tuning.

        Takeaway: transfer learning and augmentation is still effective for medium-scale datasets.

   4. [P4] Waste Classication Using Articial Intelligence Techniques A Review (2024)

      This survey covers ML and DL approaches for waste sorting. Highlights are:

      • Labeled list of datasets and open-source benchmarks.

      • Empirical guidelines for dataset splits and evaluation metrics.

      • Comparison demonstrating CNNs perform better than traditional ML approaches like SVM and RF on image- based problems.

        Its utility value comes in the form of offering dataset inven- tories and benchmarking evaluation guidelines.

   5. [P5] SWAN Platform: Industrial Symbiosis for Waste Reuse (Waste Management & Research 2021)

      This research presents a web tool to facilitate industrial symbiosis. Highlights are:

      • Purpose: Minimize information asymmetry between re- users and waste producers by leveraging structured meta- data and matchmaking algorithms.

      • Design: Central database of waste streams with attribute- and space-based decision-support services.

      • Roll-out: Piloted with regional stakeholders, illustrating how structured registries facilitate resource loops.

        The paper highlights that data infrastructure is critical to enable the translation of classication outputs to reuse trans- actions.

   6. [P6] Intelligent Solid Waste Classication Using Image Processing and Machine Learning Models (2023)

      This paper introduces a hybrid approach to automated solid waste classication by leveraging image processing methods in conjunction with machine learning and deep learning models. Major contributions are:

      • Creation of an image preprocessing pipeline through re- sizing, noise removal, and color/texture feature extraction in order to enhance classication performance.

      • Comparative analysis of state-of-the-art ML models (SVM, Random Forest, Decision Tree, kNN) with deep learning baselines (CNN).

        item Results indicated that Random Forest and SVM maintained accuracies of approximately 85

        item Allowing identication of major challenges like dataset imbalance, lighting/background sensitivity and small dataset size impacting real-world deployment.

      • Suggested directions for the future include employing larger and more varied datasets, multimodal sensing (e.g., RGB-D, spectral), and CNN model optimization for real- time smart bin and MRF usage.

        This work shows the viability of deep learning methods over traditional ML approaches to robust solid waste categorization and offers a baseline scheme for smart waste management systems.

   7. [P7] Survey and Dataset Analysis for Waste Classication (2024)

      This survey paper enumerates datasets and their character- istics that are readily available. Contributions are:

      • Exhaustive overview of datasets like TrashNet, TACO, and Kaggle Garbage.

      • Discussion of usual preprocessing operations and feature- engineering methods (HOG, color histograms).

      • Focus on dataset imbalance and necessity for multi- object, real-world data.

        It supplements SLR studies by exclusively addressing dataset availability and limitations.

   8. [P8] Case Studies of Waste Detection and Classication (2024)

      This paper documents pilot deployments of smart bins and MRF prototypes. Observations are:

      • Edge vs. cloud trade-offs: On-device inference is low latency but involves pruning and quantization. Cloud- based inference is more model-size permitting but has bandwidth expense.

      • Operational issues: Real-world issues comprise lighting variation, camera maintenance, and conveyor integration.

        These case studies illustrate the disconnect between laboratory precision and real-world system stability.

3. Comparative Analysis Across Papers

   Throughout the studies under review, the following common trends and observations can be seen:

   • Methodological variety: Articles [P1], [P4], and [P7] offer surveys and systematic reviews, whereas [P2], [P3], and [P8] demonstrate applied detection and classication systems. Articles [P5] and [P6] offer insights into indus- trial symbiosis platforms and hybrid MLDL architec- tures, respectively.

   • Trends in performance: Transfer-learning-based CNNs always deliver 8592

   • Dataset problems: All review-driven publications high- light limitations of available datasets, such as imbalance, small samples, and absence of multi-object real-world scenes. Such problems limit the capacity of models to learn from controlled lab datasets to deployment condi- tions.

   • Deployment considerations: Practitioner papers ([P5], [P8]) emphasize practical issues like lighting inconsis- tency, camera calibration, conveyor integration, and the need for structured metadata systems for facilitating circular-economy opportunities.

   • Innovation direction: The direction of future research is toward multimodal sensing (RGB + depth/spectral), larger standardized benchmark datasets, and closer inte- gration of vision systems with industrial and consumer platforms.

   Notably, for the purpose of our project we have decided to extend the work as outlined in [P5] SWAN Platform: Industrial Symbiosis for Waste Reuse. We made this choice

   because the SWAN platform focuses on the most critical gap between automated waste classication and the reutilization of materials in industrial systems. Whereas most of the other research concentrates on enhancing the accuracy of detection, the SWAN framework shows how classied waste data can be organized, disseminated, and compared with possible re-users in reality. Our project therefore aims to take this integration a step further by merging AI-based waste classication with platform-level solutions based on SWAN to ensure technical precision and real-world relevance.

   sectionMethodological Synthesis labelsec:methods

   This part generalizes frequent algorithmic and engineering themes from the eight studies and places them in relation to frequent research options.

   1. Algorithmic Families

      Broadly speaking, methods are categorized into four types:

      1. Classical ML (feature-based).: Commonly employed baselines: HOG, LBP, color histograms with SVM, RF, KNN classiers. Strengths: low compute, more interpretable. Weak- nesses: poorer performance on hard visual variability.

      2. CNN Classiers (transfer learning).: Most contem- porary classication workows utilize ImageNet-pretrained backbones (ResNet, DenseNet, MobileNet) and ne-tune on domain data. ransfer learning provides strong gains in low- to-moderate data regimes and is insensitive with standard augmentations.

      3. One-stage detectors (YOLO family).: For detection and localization operations required in MRFs and robotic pickers, one-stage detectors (YOLOv3/v5/v8) are preferred due to their provision of bounding boxes and high throughput. Adding attention modules (e.g., CBAM) and feature pyramid renements enhances small-object recall.

      4. Hybrid and multimodal systems.: Merging RGB with depth, NIR or spectral sensors adds discriminative power. For instance, depth separates items stacked or overlapping, and spectral/NIR separates material composition (e.g., glass vs. plastic) that is visually similar.

   2. Preprocessing and Augmentation

      Shared preprocessing/augmentation steps across papers:

      • Resize and normalization to match model input size.

      • Geometric augmentations: random ips/rotations, scale jitter.

      • Photometric augmentations: brightness, contrast, color jitter.

      • High-level detector augmentations: Mosaic, MixUp, ran- dom erasing; synthetic overlays for simulating occlusion and conveyor blur.

      • Class-balancing augmentations: oversampling, targeted synthetic generation.

   3. Training and Evaluation Protocols

   Best practices observed:

   • Use stratied splits when datasets are class-imbalanced.

   • Report both class-per metrics and total scores (accuracy, mAP@0.5).

   • For deployment papers, report latency and resource

   Table II: Representative reported performance taken from the re- viewed studies

   Study Model / Task Reported Metric

   (memory/FLOPS) metrics.

4. Dataset Characteristics and Limitations

   The works under review rely on private and public datasets combined. Table I reports representative properties.

   Table I: Representative datasets used or mentioned throughout the reviewed studies

   Dataset Classes Notes / Size

   TrashNet and derivatives 68 Public baseline; few thou-

   P2 (Arishi) YOLOv8 + CBAM

   (detection)

   P3 (Khetarpal et al.) CNN classier (transfer-learned)

   P1 (SLR) Survey/meta-

   analysis

   P6 (OptiFit-like) Hybrid 2D+3D re- construction

   mAP 89.5% on 17-class household dataset

   Accuracy 89.9% (cross- val)

   DL accuracies tend to 80 98% on carefully curated datasets; domain gap men- tioned

   Dimension accuracies: 95%+ for length/width in lab models

   Custom household dataset (P2)

   sand images; single-object images typical.

   17 3,700+ images; multi-object scenes, occlusions present.

   • Small-object & occlusion sensitivity: Detectors may suffer reduced recall on small or heavily occluded objects. Attention modules and higher-resolution inputs help to

     Local lab datasets (P3/P8) 612 Diverse sizes; small datasets

     curated for special experi- ments.

     alleviate this to some extent.

   • Multi-modal gains: Depth/spectral signals provably en- hance classication in borderline examples, but at the

   Industrial metadata (P5 SWAN)

   many Structured text entries; not image-based but allows matchmaking.

   expense of increased hardware complexity.

   SLR-curated dataset list (P1) 15+ List of small datasets show-

   ing imbalance and taxon- omy differences.

   Common limitations:

   • Imbalanced classes: There are not many labeled images for some types of waste, resulting in biased classiers.

   • No multi-object / conveyor scenarios: Most datasets were taken as clean single-object images; actual MRF scenes are much more complicated.

   • Taxonomy mismatch: Various studies employ different class names, making it difcult to perform transfer learn- ing and aggregation.

   • Domain gaps: Models performance usually suffers when models are trained in lab data and deployed to actual deployment sites.

5. Performance Analysis and Comparative

   Results

   This section sums up reported numbers from the papers and presents interpretation. Due to differing datasets and protocols, absolute comparisons need to be treated with caution; however, trends are signicant.

   1. Representative reported results

      Table II summarizes representative results reported through- out the reviewed studies.

   2. Interpretation and caveats

      • Strong lab performance: CNNs and contemporary de- tectors perform well on benchmark datasets. The gures quoted, though, are often based on controlled imaging scenarios and fail to represent the variability encountered in operational environments.

6. Applications and Deployment Scenarios

   Automated classication enables several deployment sce- narios; these are outlined below with practical implications.

   1. Smart Bins (at-source segregation)

      Smart bins use small cameras and edge inference to classify wastage and optionally sort it into internal compartments. Design considerations:

      • Edge constraints: Model optimization (pruning, quanti- zation) is needed for low-power CPUs/accelerators (e.g., Coral, Jetson Nano).

      • User UX: Proper usage is facilitated by clear instructions and feedback (LEDs, app notications).

      • Privacy: Processing on the device minimizes image transfer; anonymization of logs maintains privacy.

   2. Material Recovery Facilities (MRFs)

      MRFs need high-throughput detection and integration with mechanical sorters. Considerations:

      • Lighting and enclosure: Controlled lighting diminishes domain shift.

      • Multi-camera congurations: Offer several viewpoints for pose estimation and grasp planning.

      • Real-time constraints: High FPS detection and low- latency communication to actuators are necessary.

   3. Robotic Pickers

      Vision-guided robot pick-and-place systems need precise bounding boxes and sometimes per-pixel segmentation for grasp planning. Depth cameras assist in computing grasp affordances.

   4. Industrial Symbiosis and Marketplaces

   Marketplaces such as SWAN and comparable systems take structured representations of waste streams and match them against possible re-users. Vision systems can enrich these marketplaces by verifying material types and adding prove- nance metadata. Security requirements include secure data sharing, trustworthy metadata, and transparent business models to motivate participation.

7. Challenges, Limitations, and Non-Technical

   Barriers

   Though technical advances are considerable, there are chal- lenges technical, operational, and socio-economic that need to be tackled.

   1. Technical challenges

      • Dataset and benchmark standardization: Without stan- dard taxonomies and benchmarks, methods cannot be compared.

      • Domain adaptation: Models trained at one site perform poorly at another; unsupervised domain adaptation and test-time adaptation methods are required.

      • Small-object detection: Small pixel-footprint objects re- man difcult; attention and multi-scale feature extractors assist.

   2. Operational and deployment challenges

      • Hardware and maintenance: Camera cleaning, dusting, and misalignment are normal operational activities often underreported in academic articles.

      • Integration complexity: Connecting vision outputs to actuators and logistic processes (transport, baling, mar- ketplaces) involves systems engineering.

   3. Societal, regulatory, and economic barriers

      • Privacy and data governance: Personal data may be present in captured images; privacy-enhancing designs and data governance are essential.

      • Economic viability: Local markets for secondary mate- rials underpin recovery economics; classication alone is not a sufcient guarantee of protable reuse.

      • Regulatory frameworks: Incentivizing reuse through policy and harmonization of reporting can drive adoption forward.

8. Practical Recommendations for

   Researchers and Practitioners

   Based on the reviewed studies and wider literature, the following action points are suggested:

   1. Apply standard taxonomies: Creating a universally accepted class ontology (household, MRF, industrial categories) will enhance comparability.

   2. Curate multi-scene datasets: Publicly release datasets containing conveyor and multi-object scenes to bridge the labeld gap.

   3. Report resource metrics: Always report latency, model size, and energy consumption for deployment-critical research.

   4. Use multimodal sensing where possible: Combine RGB with depth or NIR when budget and maintenance allow.

   5. Pilot cross-site studies: Deploy and test models at several physical locations to estimate generalization and operational resilience.

9. Research Roadmap

   Planning a research roadmap is critical to govern the sys- tematic development of automated waste sorting and waste- to-resource platforms. As research matures, scientists must balance short-term practical demands with long-term sustain- ability objectives. Phased development brings clarityshort- term activities can bridge current gaps, while medium- and long-term initiatives advance toward more sophisticated and connected solutions. Such a roadmap emphasizes technical priorities while aligning with economic, environmental, and social needs.

   1. Short term (12 years)

      • Prioritize the development of balanced and standardized datasets.

      • Enhance lightweight models for smart bins and real-time classication abilities.

        These measures ll the short-term data quality and efciency gaps, making deployment in small-scale recycling facilities feasible and cost-effective.

   2. Mid term (24 years)

      • Establish resilient domain adaptation pipelines (unsuper- vised and semi-supervised).

      • Conduct federated learning pilots across municipalities to develop robust models while maintaining privacy.

      • Integrate vision with industrial metadata platforms to execute material ows.

   3. Long term (4+ years)

      • Standardize regulatory systems for sharing data and provenance in industrial symbiosis.

      • Achieve end-to-end automated sorting and marketplace integration that closes loops at scale.

10. Conclusion

In conclusion, this survey collates heterogeneous contribu- tions from research to present a comprehensive picture of automated waste categorization and waste-to-resource plat- forms. The review outlined how developments in datasets, algorithms, and industrial platforms are shaping the future of waste management. Even with remarkable advancements in deep learning and detection models, open problems such as dataset imbalance, domain adaptation, and deployment hurdles still exist. By synthesizing both technical and practical results,

This paper is supposed to assist researchers and those working in the sector discover right ways to solve troubles.

• The best current deep learning methods work well on specic datasets that have been carefully prepared, but their performance in real situations is not always as good because the data they use is not always realistic or comprehensive.

• Detectors that work in a single step, such as various versions of YOLO, along with attention mechanisms and robust data augmentations, can be used for most tasks.

• To make the results from these tools useful in practical ways, they need to be connected to larger systems, such as marketplaces or platforms like SWAN that help people reuse waste in economic ways.

• The community should focus on developing public test sets that cover many types of scenes, use common cate- gories, employ diverse data collection methods, and en- sure that training occurs in a way that preserves peoples privacy.

By following these ideas, people can move automated waste sorting from being demonstrated only in tests to being used in real, large-scale applications that strengthen the circular econ- omy. This paper aims to guide researchers and practitioners toward effective solutions.

References

1. F. Fotovvatikhah, I. Ahmedy, R. M. Noor, and M. U. Munir, A Systematic Review of AI-Based Techniques for Automated Waste Clas- sication, Sensors, vol. 25, no. 9, p. 3181, 2025.

2. A. Arishi, Real-Time Household Waste Detection and Classication for Sustainable Recycling: A Deep Learning Approach, Sustainability, vol. 17, no. 4, p. 1902, 2025.

3. D. Khetarpal, S. Tripathi, P. Khanna, and R. Mehta, Trash Detection: Advanced Classication of Waste Materials Using ML Techniques, in Proc. IEEE Int. Conf., 2024.

4. [Author(s) missing], Waste Classication Using Articial Intelligence Techniques: A Review, (unpublished, uploaded le: Waste Classica- tion Using Articial Intelligence.pdf ).

5. A. Angelis-Dimakis, A. Arampatzis, K. V. Moustakas, and D. Aido- nis, SWAN platform: A web-based tool to support the development of industrial solid waste reuse business models, Waste Management Research, vol. 39, no. 3, pp. 489498, 2021.

6. R. B. Raq, K. M. Hoque, M. A. Kabir, S. Ahmed, and C. Laird, OptiFit: Computer-Vision-Based Smartphone Application to Measure the Foot from Images and 3D Scans, Sensors, vol. 22, no. 1, 2022.

7. [Author(s) missing], Survey / dataset analysis for waste classication, (uploaded le: 1-s2.0-S2199853124000970-main.pdf ), 2024.

8. [Author(s) missing], Pilot deployments and practical notes on waste detection, (uploaded case study, supplemental le), 2021.

9. J. Redmon and A. Farhadi, YOLOv3: An Incremental Improvement,

   arXiv:1804.02767, 2018.

10. W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and A.

    C. Berg, SSD: Single Shot MultiBox Detector, in Proc. ECCV, 2016,

    pp. 2137.

11. K. He, X. Zhang, S. Ren, and J. Sun, Deep Residual Learning for Image Recognition, in Proc. CVPR, 2016, pp. 770778.

12. G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, Densely Connected Convolutional Networks, in Proc. CVPR, 2017,

    pp. 47004708.

13. K. Thung and C. Yang, TrashNet dataset, Public Dataset, 2016.