Thanks to visit codestin.com
Credit goes to www.scribd.com

0% found this document useful (0 votes)
25 views10 pages

S

The document reviews state-of-the-art waste management systems, highlighting advancements in waste classification through technologies like CNNs, IoT, and machine learning. It discusses various studies that propose automated waste segregation methods, their strengths, and weaknesses, particularly focusing on issues like limited classification granularity and unrealistic testing environments. The review identifies gaps in existing solutions, such as the lack of mobile-first design and user-centered education, which hinder effective waste management implementation.

Uploaded by

phdscholar740
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as TXT, PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
25 views10 pages

S

The document reviews state-of-the-art waste management systems, highlighting advancements in waste classification through technologies like CNNs, IoT, and machine learning. It discusses various studies that propose automated waste segregation methods, their strengths, and weaknesses, particularly focusing on issues like limited classification granularity and unrealistic testing environments. The review identifies gaps in existing solutions, such as the lack of mobile-first design and user-centered education, which hinder effective waste management implementation.

Uploaded by

phdscholar740
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as TXT, PDF, TXT or read online on Scribd
You are on page 1/ 10

State-of-The-Art

Waste management has become an increasingly pressing global issue, prompting


considerable research efforts to enhance waste sorting and recycling methods.
Several
technologies and methodologies have been investigated, ranging from traditional to
advanced
machine learning techniques. This section critically reviews the state-of-the-art
in waste
classification, outlining the merits and drawbacks of current solutions and
situating this effort
within the larger research landscape [13].
2.1 Existing Smart Waste Management Systems
2.1.1 SMART WASTE SEGREGATION USING IMAGE PROCESSING IN CNN [14]
Dr. J. Suresh and Dhivya Dharshini G.'s study ‘Smart Waste Segregation Using Image
Processing in CNN’ addresses the world's growing waste issue by presenting an
automated
waste segregation system based on Convolutional Neural Networks (CNNs) and image
processing techniques. The primary goal is to divide waste into four categories:
paper, food
waste, plastics, and metals, which will improve the efficiency and safety of waste
management
processes.
Strengths:
Automation and Efficiency: The use of CNNs for waste sorting introduces a high
level of
automation, potentially lowering the need for manual labor. This can result in more
consistent
and rapid waste sorting. CNNs are well-known for their accuracy in image
recognition tasks.
Using this feature, the system intends to achieve precise waste classification,
which is critical
for optimal recycling and disposal.
Safety Improvements and Scalability: Automating the segregation process reduces
human
exposure to hazardous waste, lowering the health risks associated with hand
sorting. The
suggested framework may be scaled and adapted to a variety of situations, from tiny
towns
to huge cities, making it a versatile solution for a wide range of waste management
requirements.
Integrated Approach: The study employs an end-to-end approach while addressing the
practicality of data input and output. Waste classification is more than just an
algorithmic
problem, it also demands data management from image collecting to result
interpretation.
The end-to-end element emphasizes the solution's practical effectiveness.
Weaknesses:
Lack of Technical Specificity: The IJCRT work lacks specific technical details on
CNN
architecture, training settings, and dataset construction, which inhibits
reproducibility and
comparative research. Without explicit layer configuration specifications (number
of
convolutional layers, filter sizes, activation functions), it is difficult for
other researchers to
reproduce the model's performance or adapt it to new waste categorization problems.
Similarly, the lack of information about the training method (batch size, learning
rate,
optimization algorithm) and dataset characteristics (number of photos per class,
image
resolution, data augmentation techniques) exacerbates the problem of repeatability.
This lack
of transparency not only affects the credibility of the claimed results, but it
also limits the
ability of future research to expand on the IJCRT paper's conclusions.
Simplistic Categories: The IJCRT paper's waste classification system's reliance on
broad
categories, such as ‘recyclable’ and ‘non-recyclable’, significantly limits its
practical usefulness
7
in real-world waste management scenarios. While this degree of categorization may
be
enough for early waste sorting, it falls short of meeting the complex needs of
effective
recycling systems. For example, the inability to discriminate between different
types of
plastics (PET, HDPE, PVC) or paper grades (newspaper, cardboard, office paper)
impairs the
system's ability to direct users to suitable disposal procedures. Furthermore, the
lack of sub-
categorization precludes the system from accounting for local recycling rules,
which
frequently differ by municipality or region.
2.1.2 Smart Waste Management and Classification Systems Using Cutting
Edge Approach [15]
Cheema et al. introduces SWMACM-CA, a real-time waste management and classification
platform that integrates IoT sensing, grid-based image segmentation, cloud-trained
deep
learning (VGG16), edge inference on a Raspberry Pi, and automatic sorting using a
robotic
gripper. They claim improved performance above prior art with an overall
classification
accuracy of more than 90%.
Strengths:
Grid-Segmentation for Complex Scenes: The technique partitions a congested rubbish
pile
into a 5x6 grid of uniform ‘cells’ effectively reducing multi-object photos into
single-object
segments, easing categorization tasks. The pragmatic method reduces processing
complexity
and aligns the test distribution with the single-object training set.
Edge-Cloud Hybrid Architecture: Training on a cloud AI module (Google AI) and
deploying the
inference model on an edge node (Raspberry Pi 4) provides a balance of accuracy and
latency.
This hybrid paradigm lowers round-trip delays and offloads intensive processing,
crucial for
real-time IoT applications.
High Reported Accuracy: Out of the three architectures studied (VGG16, Fast-RCNN,
and
MobileNetV2), VGG16 surpassed both Fast-RCNN (88%) and MobileNetV2 (85%), with an
accuracy of 96.1% on the TrashNet subset, exceeding the project's stated goal of
90%.
Weaknesses:
Controlled-Environment Bias: All studies were conducted under strictly controlled
conditions,
including uniform lighting, single-object grid cells, and clean backgrounds
inspired by
TrashNet's ‘white poster’ iconography. Such limits fail to capture the diversity of
real-world
waste yards, which feature overlapping objects, shadows, mud, and a variety of
backgrounds.
Real-world clutter challenges the grid-segmentation assumption that each cell
contains only
one object, potentially leading to misclassification or missing detections.
Limited Waste Taxonomy: The approach divides waste into four categories: plastic,
metal,
glass, and ‘trash’ leaving out paper/cardboard, organic (biodegradable) waste, and
hazardous
streams. This restricted taxonomy overlooks key municipal waste parts, limiting its
relevance
in complete recycling or composting systems.
Dataset Imbalance and Domain Shift: The authors notice the class imbalance in the
TrashNet
subset (e.g., 137 ‘trash’ photos vs. 501 glass), but they do not use balancing or
augmentation
approaches. Furthermore, the extreme contrast between curated training photos and
chaotic
dump-yard scenes creates significant domain change, potentially decreasing real-
time
performance.
Omitted Latency and Throughput Metrics: Although the hybrid architecture aims to
reduce
latency, the article lacks quantitative measurements for end-to-end inference time,
grid
8
segmentation overhead, and actuation delays. Without these, claims of ‘real-time’
operations
are baseless.
2.1.3 Waste Segregation Using Image Detection and IOT [16]
Rizwana S. et al.'s research paper ‘Waste Segregation Using Image Detection and
IoT’
describes an automated waste classification system built with YOLOv8, OpenCV,
Arduino Uno,
and Firebase. The project intends to increase waste management efficiency by
automating
segregation, which is an important step in recycling and environmental
sustainability. While
the research proposes a potential technique, several areas warrant more
investigation to
determine its validity, applicability, and scalability.
Strengths:
End to End Integration: One of the paper's most notable contributions is its
comprehensive
approach, the authors go beyond algorithmic categorization and connect the vision
model
with actual hardware (Arduino Uno + servo motor + ultrasonic sensor) and a web
dashboard.
This ‘full-stack’ mindset, encompassing data collection, inference, actuation, and
remote
monitoring, is uncommon in academic prototypes, which frequently focus on a single
slice
(e.g., the model) in isolation. By illustrating the entire pipeline from camera
input to bin
actuation, this work persuasively demonstrates the viability of implementing such
systems in
real-world settings.
Adoption of State-of-the-Art Detection Model: The usage of YOLOv8 (the most recent
iteration of the You Only Look Once family) puts the work at the forefront of real-
time object
detection. YOLOv8 outperforms previous versions (e.g., YOLOv5) in terms of accuracy
and
speed, and the authors reduce training time and data needs by exploiting transfer
learning
from the COCO dataset. The team's approach is practical, as they build on a well-
supported
framework rather than reinventing it.
User Centred Features: The system not only opens bins automatically, but also
allows users to
upload photographs or videos of waste for offline classification, expanding its
application.
Webcam-based real-time sorting and manual image upload provide a range of
deployment
scenarios, including smart bins in public locations and desktop programs for
training and
auditing waste processes.
Visualization and Monitoring: The authors' dashboard shows pie-chart and bar-graph
summaries of waste proportions, while the Firebase-backed fill-level display
provides real-
time bin capacity feedback. These visualizations improve usability for waste
management
operators by supporting data-driven collection schedules and identifying trends in
trash
composition.
Weaknesses:
Lack of Quantitative Evaluation Metrics: The study does not include common
performance
indicators like accuracy, precision, recall, F-score, mean average precision (mAP),
or inference
latency. Instead, it uses screenshots of accurate classifications. The anecdotal
evidence does
not show how the model performs in edge instances, unusual classes, or with real-
world
unpredictability. Furthermore, there is no comparison with baseline procedures
(e.g.,
traditional image-processing pipelines or prior YOLO versions), making it unclear
if YOLOv8
delivers a meaningful improvement in this scenario.
Limited Classification Taxonomy: The four basic categories, wet, dry, metal, and
plastic,
represent a simplified picture of urban waste streams. In practice, additional
kinds (glass,
paper, hazardous/e-waste, compostables) are essential. The authors discuss future
expansion
but do not illustrate the system's extensibility. Multi-material objects, such as
plastic-metal
9
composites, present a classification difficulty that the current taxonomy does not
adequately
handle.
No Field Trials or User Studies: All outcomes appear to be limited to laboratory or
virtual
environments. The system's adoption by waste management workers and end-users
(citizens
depositing rubbish) is untested. Human factors, such as how people react to
automated bin
lids or interact with the dashboard, are ignored.
2.1.4 Automated Waste Segregation System Using Opencv And Image
Processing [17]
Kavade et al. propose an automated system that uses computer vision (OpenCV + YOLO
v5),
simple actuators (servo-driven flaps), and an ESP32-CAM module to sort items into
bins,
addressing the issue of manual waste segregation that leads to environmental
contamination
and biohazards. This is current and consistent with the broader ‘smart city’ ideas.
However,
the study fails to place its contribution inside a quantifiable performance gap,
there is no
explicit description of the accuracy, latency, or cost constraints of existing
systems that this
work seeks to overcome.
Strengths:
Integration of Multiple Disciplines into a Working Prototype: This research
successfully
combined computer vision, embedded systems, mechanical design, and basic
mechatronics
to create a workable prototype. Many studies in this field concentrate only on the
software
or conceptual modeling of the waste segregation challenge. However, Kavade et al.
go beyond
theoretical design and create a practical, functional prototype that combines real-
time item
detection, classification, and mechanical actuation. The authors use the ESP32-CAM
module,
Arduino Uno, and servo motors to demonstrate their understanding of embedded system
restrictions, sensor interfacing, and actuator control. This makes the system both
conceptually
plausible and technically demonstrable. Such integration helps to validate their
approach in
practice, rather than merely in simulation.
Appropriate Choice of Object Detection Model (YOLOv5): The use of YOLOv5 as the
primary
object detection framework is also interesting. YOLOv5 is well-known for striking a
compromise between speed and precision, making it ideal for real-time applications
on
resource-constrained devices. It’s lightweight architecture, as compared to
previous versions
and models such as R-CNN or SSD, enables faster inference without sacrificing
classification
performance. The authors' choice of YOLOv5 over heavier models or less accurate
classifiers
demonstrates an understanding of the trade-offs between real-time performance and
computational overhead, particularly given that the target deployment includes low-
cost,
low-power devices such as the ESP32-CAM and Arduino microcontrollers.
Sustainable and Scalable Design Considerations: Another advantage is the
incorporation of
scalability and environmental sensitivity into the design. The authors suggest a
design that
might be powered by solar panels, encouraging the use of clean energy in urban
applications.
Furthermore, the mechanical design is meant to be modular and adaptable to many
types of
dustbins and settings. This makes the system appealing for deployment in semi-
urban, rural,
or developing locations where conventional systems may be prohibitively expensive
or lack
adequate infrastructure. By including portability, solar power, and adaptability,
the study fits
well into the landscape of sustainable smart-city developments.
Weaknesses:
Lack of Quantitative Metrics and Performance Benchmarks: The most significant issue
is a lack
of comprehensive quantitative examination. The authors consistently claim ‘high
accuracy’ in
10
waste classification, but fail to back this up with empirical measures such as
precision, recall,
F1-score, confusion matrices, mean average precision (mAP), inference time (delay
per
frame), and processing throughput (items per second or minute). Without this, it is
impossible
to objectively evaluate the system's categorization reliability.
Vague Dataset Details and Training Procedure: The model's training procedure is
insufficiently
documented, which greatly limits reproducibility. The authors state that they
employed ‘non-
real-time images’ for initial training before refining it with real-time data. This
opacity in
dataset details indicates either an insufficient experimental technique or a lapse
in scientific
reporting. Furthermore, it is unknown whether the dataset was gathered by the
authors or
obtained from public sources such as TrashNet, and how well the training data
reflects real-
world situations (e.g., dirty, damaged, or partially visible objects).
Idealized Testing Scenarios: All of the test findings and photographs included in
the
publication indicate that the system was tested under controlled, static
conditions, single-
item placement, neutral backdrops, and adequate illumination. This is not
representative of
real world surroundings, where several waste items are frequently present, things
may be
stacked, crushed, or obstructed, lighting may be uneven (indoors, outdoors,
nighttime), and
items are damp, unclean, or broken. There is no testing in these complex, congested
situations, reducing trust in the system's robustness. Systems intended for real-
world
deployment must be tested under stress conditions, such as overlapping waste kinds
and
deformable or reflecting surfaces.
2.2 Identified Gaps in Existing Solutions
Limited Waste Classification Granularity
Most existing systems categorize waste into limited broad categories, such as
recyclable, non-
recyclable, organic, or dry/wet distinctions. While this provides a high-level
insight, it lacks the
detail needed for precise waste management. For example, recycling protocols differ
between
plastics (e.g., PET vs. PVC), metals (aluminum vs. tin), and papers (shredded vs.
waxed).
Systems such as those developed by Rizwana et al. and Suresh & Dharshini have only
four
categories, making them inappropriate for regions with complex waste sorting rules.
A more
defined taxonomy is critical for municipalities, recycling centers, and
environmental
compliance.
Unrealistic Testing Environments
A common concern in academic prototypes is the use of clean, idealized datasets
(such as
TrashNet or self-curated studio-quality photos) for model training and evaluation.
This ignores
frequent environmental issues such as inadequate lighting, occlusion, cluttered
backdrops,
damaged goods, and various object orientations. As a result, these systems are not
ready for
use in the field and have difficulty generalizing. For example, a takeaway
container covered
with food residue may not be properly designated as recyclable. This distance from
the
complexities of real-world waste hampers effective implementation.
Absence of Mobile-First Design
While a few studies use embedded hardware (e.g., Raspberry Pi, Arduino, ESP32),
these
configurations are specialized, immobile, and expensive to scale. Others use cloud-
based
processing, which is both latency-prone and privacy-sensitive. None take advantage
of
cellphones' widespread availability and computational capability to provide
sophisticated
waste management capabilities to everyday users. This is a serious mistake, given
that over
11
6.9 billion people worldwide hold mobile phones, which provide an unprecedented
potential
to democratize sustainable behavior.
Lack of User-Centered Education and Feedback
The great majority of current systems function as passive classifiers, providing no
post-
classification education, motivation, or individualized feedback. In actuality,
waste sorting
behavior is extremely behavioral, users benefit from clear, contextual information
about why
an item belongs in a specific bin and how to improve their performance over time.
Current
systems lack behavioral change models, instructional nudges, and localized policy
information, resulting in low user engagement and learning.
Incomplete System Integration
While some projects use hardware actuation (for example, servo-controlled bins),
few provide
a completely integrated software experience. There is a lack of permanent tracking,
real-time
dashboards, and offline capabilities. Without lasting user data or localized
intelligence, these
products are only short-term fixes rather than components of a long-term waste
literacy
strategy.
No Standardized Evaluation Metrics
The usage of defined benchmarks is critical to the development of any intelligent
system,
mean Average Precision (mAP), latency (ms), throughput (fps), false
positive/negative rates,
and so on. Many reviewed systems fail to report such metrics or do so
inconsistently, making
meaningful comparison and replication impossible. Furthermore, few usability tests
are
conducted to evaluate the human-computer interaction (HCI) aspect of the solution.
2.3 Positioning of This Project
This project was created in response to a well-defined set of constraints found in
the present
landscape of intelligent waste management solutions. Rather than treating the
problem solely
from a technical aspect, the project was designed to bridge the gap between
theoretical study
and practical, user-centered implementation. The system architecture,
functionality, and
assessment process were all influenced by a critical understanding of the strengths
and
limitations revealed in previous work.
2.3.1 Bridging the Gap in Waste Classification Granularity
One of the most common problems discovered in previous research is simplicity of
waste
categorization. While many systems restrict themselves to basic classifications
like ‘recyclable’
vs. ‘non-recyclable’ or ‘organic’ vs. ‘dry’ this frequently results in
misclassification and
misunderstanding in real-world disposal settings. For example, different types of
plastics or
composite materials have separate recycling methods, and combining them under a
single
label reduces the accuracy required for sustainable waste management [18].
This project starts with a simple binary classification, differentiating between
recyclable and
organic materials, but is architecturally built to expand to more granular classes.
The dataset
selection, data pipeline, and training framework were designed to allow for future
enhancements, such as identifying PET, HDPE, glass, aluminum, and e-waste
independently.
This extensibility is an intentional design decision that directly addresses the
limitations of
existing systems, which lack the ability to evolve with more sophisticated sorting
procedures
or region-specific waste regulations.
12
2.3.2 Grounding the System in Realistic Testing Conditions
Another major issue in much of the related literature is the use of clean,
laboratory-style
datasets for training and testing. These databases, which are frequently made up of
idealized
photographs, fail to capture the common complexities of waste presentation, such as
crumpled materials, unclean containers, and uneven lighting conditions. This
difference
creates a significant barrier to real-world adoption, as models trained on cleaned
inputs
struggle to generalize.
This study involved training and testing on a more diversified dataset that
included noisy,
flawed, and contextually rich images. The training approach used picture
augmentation
techniques to recreate realistic variations, allowing the model to acquire robust
features that
are unaffected by background clutter, lighting differences, and partial occlusion.
As a result,
the system exhibits great generalization capacity, progressing beyond academic
novelty to
field readiness.
2.3.3 Implementing Mobile-Centric Design for Accessibility
Many existing solutions are either dependent on custom hardware, which can be
expensive
and difficult to scale, or on cloud infrastructure, which introduces latency,
reliance, and
privacy concerns. This project takes a deliberate step away from existing
techniques by
implementing a mobile-first architecture. Recognizing that smartphones are among
the most
widely used computing devices in the world, this research takes advantage of their
embedded
computational capability to execute inference locally and offer immediate results.
By adapting the VGG16 model for mobile compatibility and efficiency, the app allows
waste
classification to take place directly on the user's device. This reduces the
requirement for
constant internet access, making the solution suitable for rural or bandwidth-
constrained
places. Furthermore, implementing on mobile platforms significantly boosts
scalability and
potential impact, making smart waste guidance available to a wider audience without
the
need for specialist equipment.
2.3.4 Integrating Educational Feedback Through AI-Powered Support
While most existing systems finish their obligation at classification, this project
takes a more
instructional and behavior-oriented approach. The inclusion of AI as a real-time
assistant
enables the system to do more than just classify stuff. It actively explains the
logic for each
classification, provides extra disposal advice, and responds to concerns about area
waste
policies or material breakdown.
This interaction layer is more than just a value-added feature; it tackles a
fundamental flaw in
existing solutions: a lack of user involvement and behavioral direction. Waste
sorting is a
behavior-driven process, and users greatly benefit from contextualized instruction.
Through
AI integration, the system promotes learning and continual development, promoting
itself as
a platform for long-term behavior modification rather than just a tool.
2.3.5 Building Toward Complete System Integration
Many prototype systems have a noticeable shortcoming in that they are short-term or
one-
dimensional. Some rely only on hardware automation (e.g., smart bins), while others
offer
one-time mobile classifiers with no persistent capability. In contrast, this
project was designed
with a bigger vision in mind, emphasizing long-term usability, modularity, and
extensibility.
The present program consists of modules for classification, feedback transmission,
user
interaction, and performance logging. It enables local caching and offline use,
allowing it to
function under a variety of network situations. Furthermore, its modular
architecture allows
13
for future additions such as IoT bin connection, personal consumption dashboards,
and long-
term waste tracking to be implemented with minimal system changes. This integrated
approach frames the initiative as a useful component of bigger smart city
ecosystems and
waste awareness campaigns.
2.3.6 Promoting Transparent and Reproducible Evaluation
This project's evaluation process differs from that of many previous systems, which
report
performance metrics inconsistently or omit them entirely. Chapter 5 explicitly
defines and
presents a comprehensive array of measurements, including accuracy, precision,
recall, F1-
score, mean Average Precision (mAP), latency, and throughput. These measurements
ensure
that the model's performance is rigorously evaluated and publicly published, so
that future
researchers and developers can confidently build on this work.
Furthermore, the project provides visual evaluation tools such as confusion
matrices, ROC
curves, and confidence histograms, which provide a better understanding of the
classifier's
strengths and edge situations. This emphasis on transparency and reproducibility
addresses a
significant gap in the current literature, which frequently includes ambiguous or
unverifiable
statements.
Project Impact
This project represents a strategic evolution rather than a reinvention of previous
work. It
builds on current research while purposefully addressing its limitations, combining
mobile
accessibility, real-world generalization, instructional AI integration, and
transparent
evaluation to create a unified and deployable waste classification system.
Each design decision was motivated by the need to close existing gaps, ranging from
fine-
grained classification potential and user behavior reinforcement to technical
robustness and
long-term scalability. By presenting the idea in this manner, it not only addresses
current limits
but also provides the groundwork for future advancements in intelligent waste
management

You might also like