Waste material classification using performance evaluation of deep learning models

1 Introduction

Waste material classification is sorted by composition, substance, and recycling or disposal possibilities. Organic, recyclable, hazardous, and non-recyclable trash are the primary waste categories. Organic waste includes waste from food, the yard, and other biodegradable sources. Paper, plastic, glass, and metal may be recycled into new items. Batteries, chemicals, and medical waste can harm humans, animals, and the environment. Plastics, foam, and dirty paper are non-recyclable. The separating garbage process should be followed by local recycling and disposal rules. Private waste management firms and towns offer curbside recycling and hazardous waste disposal. Sorting and disposing of garbage appropriately reduce landfill waste and environmental effects. Waste classification and disposal may preserve resources, minimize pollution, and save energy [1].

Identifying that most major cities’ waste is recyclable, it is important to learn and employ reuse strategies that can improve the environment. Sorting rubbish has become crucial to proper disposal. Even though there are many recycling categories, consumers might still become confused about which trash bin is appropriate for each rubbish. This research suggested the use of a deep learning (DL) models-based automated method to sort home waste into recycling categories to reduce the environmental effect of improper waste disposal. Individuals have controlled solid waste by eliminating it for ages. Population expansion drives rubbish production. To maintain a waste management balance, reduce it personally. Waste management and sorting are crucial for environmentally sustainable development worldwide. Recycling and reusing garbage is crucial for society [2].

Population growth has resulted in a corresponding rise in solid waste output. The economy, public health, and ecology are all negatively impacted by improper garbage management. It has been a problem for rapidly expanding urban centers all over the world to effectively manage their solid waste. The economic and environmental benefits of efficient trash recycling are substantial. Recovering raw materials, conserving energy, lowering greenhouse gas emissions and water pollution, decreasing the number of new landfills, etc., are all possible benefits. To achieve profit, scavengers and collectors in impoverished countries sort recyclables from municipal solid waste (MSW) at homes and trade them on the black market. More people participate in recycling programs in industrialized nations. Automatic garbage sorting may be achieved using a variety of methods including mechanical sorting and chemical sorting [3].

With the evolution of DL, incredible advances have been made in many fields, including image classification [4]. An essential in recycling and other forms of garbage management is trash classification [5]. Identifying improved methods to organize solid waste is essential due to the fact that it is being produced at a rapid pace all through the globe. It presents an advanced DL image classification model for garbage sorting, training it on a dataset divided into cardboard, glass, metal, and rubbish [6]. This study aimed to correctly categorize images of trash into these groups; doing so would have far-reaching effects on how waste is managed and the natural world. This study describes it in-depth, outlining its architecture and detailing the outcomes of our performance analysis. DL models can potentially improve waste categorization efficiency, which is crucial for the development of waste management. The key innovations and contributions can be summarized as follows:

• The study determines the most crucial features and parameters to accurately classify waste materials using DL models (Resnet50, GoogleNet, InceptionV3, and Xception). This study chooses the sensitivity analysis of feature importance to ascertain which variables significantly affect classification precision.

• To assess whether DL models can be successfully used for the problem of garbage classification in the real world, a comparison of the models in terms of their efficiency, scalability, and cost-effectiveness has been carried out.

• To evaluate the efficacy of several DL models (ResNet50, GoogleNet, InceptionV3, and Xception) for sorting waste, the models’ performance using metrics such as F1 score, accuracy, recall, and precision has been compared.

The organization of this study is as follows: the related studies on waste material classification using DL models are discussed in Section 2. Section 3 presents the proposed waste material classification using performance evaluation of DL models. The results and discussion of the proposed waste material classification using performance evaluation of DL models and their benchmarking are presented in Section 4. Finally, the conclusion and future work are presented in Section 5.

2 Related works

Garbage disposal is a severe problem that affects the environment in both wealthy and developing nations. It is recognized as the single significant threat to long-term prosperity and environmental stability. Since they are determined to serve the public interest, many NGOs are advocating for this cause and directing the attention of government agencies to this cause. The World Wide Fund for Nature is the most prominent organization in this field (outside of NGOs) [1,3]. Groups of intellectuals and environmentalists exist, with members driving their causes in their local communities.

Garbage disposal has become a major issue in recent years due to the widespread use of single-use products [1]. This includes everything from bottled water to disposable coffee cups, from packaging foam to medical waste, and from light bulbs to plastic bags. The urgent need to restore ecological equilibrium, which has been severely disrupted by human activity in the past 200 years only, is the driving force behind effective waste management.

According to Majchrowska et al. [7], over 10 trash datasets are analyzed, and previous DL-based algorithms are reviewed to solve the issues with automatic waste detection. The authors suggest updating the existing trash detection and classification benchmark datasets to include all types of garbage. By combining EfficientDet-D2’s litter localization with EfficientNet-B2’s trash classification, they describe a two-stage detector that can achieve an average precision of 70% in waste localization and an accuracy of roughly 75% in classification. The study’s code and annotations are open-source, and the suggested method is trained semi-supervised on unlabeled photos. This study establishes a repeatable standard for identifying and categorizing different types of trash, which might be used to enhance the effectiveness of waste management and lessen the negative effects of trash on the environment.

In another study by Shi et al. [8], the fast progress of DL technology has led to the proposal of several network models for classification, all of which contribute to the successful implementation of intelligent garbage categorization. However, current models in garbage categorization still have certain issues, including low classification accuracy and lengthy running times. This research proposes a waste categorization approach using a multiple-layer hybrids convolution neural network (CNN) to address these issues. The network topology of this approach is remarkably equivalent to that of VGGNet; however, it has fewer parameters and better classification accuracy as a result. The efficiency of the suggested model may be modified by adjusting the total number of network models and channels. This research selects the most efficient model for waste image classification after determining the ideal parameters for doing so. The experimental outcomes demonstrate that the suggested technique has a more straightforward network topology and improved waste categorization accuracy in comparison to some current efforts. The impact of the suggested technique is demonstrated by the results of several tests conducted on the TrashNet dataset, which shows the possibility to achieve an accuracy rate for classification of up to 92.6%, which is 4.18 and 4.6% points greater than the results of several state-of-the-art methods.

Recycling and landfilling are two methods used in the waste management process that lead to the destruction of trash [9]. Together, DL and Internet of Things (IoT) provide a quick answer to the problems of categorization and real-time data monitoring. Incorporating DL and the IoT, this article depicts a robust framework for trash management. CNNs are a common method in DL, and the suggested model uses them to intelligently separate biodegradable from non-biodegradable garbage. The approach includes an architectural outline for a microcontroller-powered, sensor-laden “smart trash can.” The suggested technique uses the IoT and Bluetooth to monitor the devices used. With the IoT, data are managed in real-time with the aid of Bluetooth for monitoring devices from a short distance using an Android app. Accuracy in waste label categorization, sensors data estimate, and system usability scale (SUS) are tallied and interpreted as means to evaluate the constructed model. The suggested architecture employing the CNN model achieves 95.3125% accuracy in classification and 86% accuracy in SUS scoring [9].

Waste materials such as glass, paper, food, paper boxes, etc., are readily accessible. Classifying waste things using computer vision-based technologies is a cost-effective method of sorting through a large rubbish dump [10]. Recent advances in DL and deep reinforcement learning (DRL) make it feasible to detect and identify wastes, allowing for their categorization into specific categories. This research develops a model for intelligent DRL-based detection and classification of recyclable waste objects (IDRL-RWODC). The IDRL-RWODC method employs DL and DRL to identify and categorize trash items. Objects are first detected using a Mask Regional Convolutional Neural Network (Mask RCNN) and then classified using DRL. In addition, a deep Q-learning network (DQLN) is used as a classifier, and the DenseNet model is deployed as a baseline model for the Mask RCNN model. In addition, a hyperparameter optimizer built on the dragonfly algorithm is devised to boost the performance of the DenseNet model. Extensive simulations were run on a benchmark dataset, and the experimental findings showed that the IDRL-RWODC strategy outperformed more contemporary methods with a maximum accuracy of 0.993 in trash classification [10].

In another study by Chaturvedi et al. [11], the dumping of MSW in urban areas has become a serious problem that, if left unchecked, might lead to environmental damage and even human health risks. To effectively handle various types of trash, a well-thought-out waste management system is essential. The primary objective of this study is to develop an advanced modeling approach that could reliably foretell the quantity of MSW that will be generated in the future. To achieve this goal, a framework was built for preliminary processing and data integration using a CNN and an air-jet system. The findings demonstrated the efficacy of maintaining machine learning algorithms for garbage classification. Artificial neural network models performed most beneficially, accounting for 72% of the variance in the data. This study’s approach indicates the viability of creating tools to aid in the management of urban garbage by supplying, preprocessing, integrating, and modeling data that are freely available from a wide range of sources.

In the study by Rutqvist et al. [12], the authors suggest utilizing machine learning to identify empty recycling bins using binary classification. The model’s superior performance can be attributed to the thorough application of data preparation procedures. By using a feature engineering strategy to extend the model, it improves accuracy over more conventional methods. Performance was optimized for the feature engineering model, traditional machine learning with basic features, and enhanced features.

In the study by Dubey et al. [13], the system provides an IoT- and ML-powered “smart bin” for household garbage collection and sorting. The authors offer an ML-based approach to manage waste. The suggested method for MSW sorting and classification is predicated on a K-Nearest Neighbor (KNN) model. With the suggested approach, organic and inorganic wastes are separated at the household level, and then further subclassified at the societal level, using a pipeline design with segmented modules. With an accuracy of 93.3%, the KNN model is superior to the alternatives.

Many studies attempt to provide solutions with different techniques to solve waste trash classification, such as [14–21]. Since trash accumulation directly impacts local citizens’ quality of life and health, it is prioritized in city planning and development. Researchers have devoted an effort to solve the challenge of automatic litter categorization. Recently, DL models have attained accuracy equivalent to that of humans, which has sparked a revolution in the field. Modern DL models are compelling, which may outperform humans at some specialized tasks.

3 The proposed waste material classification using DL models

The requirement for effective and environmentally friendly trash management highlights the significance of comparing the efficiency of different DL techniques for waste material classification. Recycling, lowering pollution levels, and preserving resources all depend on correct garbage sorting. However, to determine which DL model is appropriate in real-world scenarios, their efficacy and efficiency are compared and analyzed. To educate on recycling and waste management industry decision-making, this study sheds light on the efficacy of several models based on DL and their applicability in real-world contexts. Evaluating DL models for garbage classification is an important field of study since accurate waste classification may help create an economically viable and environmentally sensitive future. The proposed waste material classification using DL models is shown in Figure 1.

Figure 1

The proposed waste material classification using DL models [1].

4 Validation and testing results

To assess the efficacy of a model based on DL for garbage categorization, first step is to undergo validation and testing. In validation, test the model performance on a set of data that was not utilized during training. Then, the model is evaluated to observe the effect it performs in real-world settings by application on data that have not been observed before. Accuracy, precision, recall, and F1 score are measures that assess the efficacy of a DL model. Precision and recall assess the classification under study, whereas accuracy assesses the proportion of properly labeled samples. The F1 score put into consideration both accuracy and recall, weighing both equally. To evaluate the model based on DL for garbage categorization to perform its assigned function, it is validated and tested. In most cases, significant accuracy and F1 score indicate a more successful model. It is critical to test the model’s ability to generalize to novel data and prevent it from overfitting the training set. Results from validating and testing trash categorization models built using DL may be used to refine the models and zero in problem regions. These findings may also be used to evaluate other models and choose the best option without a doubt for rubbish categorization tasks.

To evaluate the methodology of the trained models’ generalization skills, validation sets are utilized. Four pre-trained models, including GoogleNet, Inception v3, Xception, and ResNet50, are used for training, with a 70/30 training/testing split on the dataset. Models are trained on the training set, then their parameters and hyperparameters are tuned using the validation set, and finally, the model’s generalization skills are assessed with the testing set. Using validation sets, overfitting tendencies in training models are identified and addressed, leading to models with improved reliability and accuracy when predicting the waste item class. Using validation sets in model development and assessment is a critical action for sustainable waste management –management efficiency. It also highlights the significance of using validation sets in DL applications to increase the adaptation characteristics of the models and overall performance. Averaged confusion matrix of four DL models running individually on 435 test images are presented.

4.1 ResNet50 evaluation results

As illustrated in Tables 3 and 4 and Figures 2 and 3, it can be perceived that the ResNet50 model achieved an impressive overall performance in classifying four classes (cardboard, glass, metal, and trash) with an accuracy of 0.95, precision of 0.954, recall of 0.95, and F1 score of 0.948. These results indicate the model’s ability to effectively differentiate between classes, with low rates of false positives and false negatives. However, within the glass category, two misclassifications were found: one sample was incorrectly labeled as metal, and the other as garbage. Potential approaches to addressing these misclassifications include gathering more labeled data, adjusting the model, and investigating alternate architectures. These enhancements, when implemented, will boost the model’s performance, leading to more precise and trustworthy predictions in practical settings.

Table 3

Averaged confusion matrix of ResNet50 running individually on 435 test images

Actual | Predicted Class	Cardboard	Glass	Metal	Trash
Cardboard	0.978	0.008	0.006	0.006
Glass	0.003	0.952	0.035	0.008
Metal	0.002	0.021	0.962	0.013
Trash	0.028	0.003	0.044	0.924

Table 4

ResNet performance metrics

Accuracy	Precision	Recall	F1
0.95	0.9545	0.95	0.9484

Figure 2

Left: confusion matrix for ResNet50; right: ROC characteristics for ResNet50.

Figure 3

Resnet50 training/validation graph.

4.2 InceptionV3 evaluation results

Cardboard, glass, metal, and rubbish were the four classes used to train the InceptionV3 model. Accuracy, precision, recall, and F1 score all came in at a flawless value of 1.0. Consequently, the outcomes were acceptable. All samples were accurately sorted into their respective classes, as shown by the confusion matrix in Tables 5 and 6 and Figures 4 and 5. There were no false positives or false negatives. Due to its advanced design, the InceptionV3 model can extract complex and discriminative characteristics from the input data, which contributes to its outstanding performance. The 100% accuracy in categorizing the four classes demonstrates the model’s dependability and resilience. The results show that InceptionV3 has the potential to be a useful approach for complex classification issues in a wide range of contexts. The generalizability of the model has to be verified in the next research by analyzing new data or using cross-validation methods.

Table 5

Mean confusion matrix from 435 test images processed by InceptionV3 running alone

Actual | Predicted Class	Cardboard	Glass	Metal	Trash
Cardboard	0.9591	0.0186	0.0104	0.0112
Glass	0.0004	0.9228	0.0698	0.0069
Metal	0.0054	0.0321	0.9173	0.0450
Trash	0.0115	0.0051	0.0346	0.9485

Table 6

Inception performance metrics

Accuracy	Precision	Recall	F1
1.0	1.0	1.0	1.0

Figure 4

Left: confusion matrix for InceptionV3 and right: ROC characteristics for InceptionV3.

Figure 5

InceptionV3 training/validation graph.

4.3 Xception evaluation results

The Xception model was also trained using the same dataset, which included four categories: paper, plastic, metal, and rubbish. As can be seen in Tables 7 and 8 and Figures 6 and 7, the outcome of the confusion matrix provides some intriguing new information. For both the metal and cardboard classes, the model obtained flawless classification, with zero misclassifications. However, the glass class exhibited slight challenges, with one misclassification as metal and another as trash. Similarly, the trash class experienced one misclassification as glass and another as metal. Overall, the Xception model demonstrated high accuracy and precision, on the other hand its performance was slightly affected by misclassifications in the glass and trash classes. These results suggest that the Xception model has the potential to effectively discriminate between classes, but it may benefit from further optimization to improve its ability to differentiate between glass, metal, and trash samples. Possible strategies include fine-tuning the model, exploring alternative data augmentation techniques, or considering ensemble approaches. By addressing these mis-classifications, the Xception model could achieve even higher accuracy and precision, making it a valuable tool for multi-class classification tasks involving cardboard, glass, metal, and trash categories.

Table 7

Averaged confusion matrix of Xception running individually on 435 test images

Actual | Predicted Class	Cardboard	Glass	Metal	Trash
Cardboard	0.9165	0.0745	0.0052	0.0036
Glass	0.0049	0.9087	0.0774	0.0088
Metal	0.0098	0.0352	0.8797	0.0752
Trash	0.0306	0.0108	0.0393	0.9191

Table 8

Xception performance metrics

Accuracy	Precision	Recall	F1
0.90	0.9078	0.90	0.8981

Figure 6

Left: confusion matrix for Xception and right: ROC characteristics for Xception.

Figure 7

Xception training/validation graph.

4.4 GoogleNet evaluation results

As seen in Tables 9 and 10, the GoogleNet model underwent training on the dataset comprising four distinct classes: cardboard, glass, metal, and trash. With an accuracy of 0.9, precision of 0.9078, recall of 0.9, and F1 score of 0.8981, the model displayed commendable performance. Analyzing the accompanying confusion matrix in Figure 8 unraveled intriguing insights. Evidently, the model exhibited impeccable precision and recall for the cardboard and metal classes, achieving flawless classification without any instances of false positives or false negatives. Nonetheless, the glass and trash classes posed challenges, as one glass sample was misclassified as metal, while one trash sample was misjudged as glass. These discrepancies exerted a noticeable impact on the precision and recall measures for these classes. Despite these isolated instances, the GoogleNet model demonstrated an admirable overall performance, characterized by substantial accuracy and precision values. To further augment its efficacy, fine-tuning the model parameters or delving into techniques like transfer learning holds potential as shown in Figure 9. By mitigating the misclassifications witnessed in the glass and trash classes, the GoogleNet model can potentially attain even greater accuracy and precision, thereby amplifying its utility for multi-class classification scenarios encompassing cardboard, glass, metal, and trash classifications.

Table 9

Averaged confusion matrix of GoogleNet running individually on 435 test images

Actual | Predicted Class	Cardboard	Glass	Metal	Trash
Cardboard	0.949	0.035	0.010	0.005
Glass	0.001	0.894	0.092	0.011
Metal	0.006	0.042	0.871	0.079
Trash	0.023	0.015	0.062	0.898

Table 10

GoogleNet performance metrics

Accuracy	Precision	Recall	F1
0.90	0.9078	0.90	0.8981

Figure 8

Left: Confusion matrix for GoogleNet and Right: ROC characteristics for GoogleNet.

Figure 9

GoogleNet training/validation graph.

5 Conclusion

The conducted analysis involved training and evaluating four different models, namely, ResNet50, InceptionV3, Xception, and GoogleNet, on a dataset consisting of four classes: cardboard, glass, metal, and trash. Each model demonstrated varying levels of performance in terms of accuracy, precision, recall, and F1 score, providing valuable insights into their effectiveness for multi-class classification tasks. ResNet50 exhibited high accuracy (0.95) and precision (0.954), along with a commendable recall (0.95) and F1 score (0.948). The model showcased reliable performance with only two misclassifications in the glass class, where one sample was mislabeled as metal and another as trash. InceptionV3, on the other hand, achieved remarkable accuracy, precision, recall, and F1 score of 1.0. The model presented adequate classification across all classes, with no misclassifications. This exemplary performance can be attributed to the model’s intricate architecture, allowing it to capture complex patterns and hierarchies within the data. Despite its impressive accuracy of 1.0, Xception has some trouble with the glass and rubbish classes. Misclassifications were seen in these subsets, leading to a decline in accuracy and recall. The model’s performance on a no-errors-classification job utilizing cardboard and metal was, on the other hand, outstanding. With an F1 score of 0.8981, 0.9078 precision, 0.90 recall, and 0.90 accuracy, the GoogleNet model did quite well. Misclassifications in the glass and rubbish classes occurred, just like in Xception. However, the model excelled in accurately classifying the cardboard and metal classes, achieving flawless precision and recall for these categories. These results highlight the strengths and weaknesses of each model. While ResNet50, InceptionV3, and Xception demonstrated overall strong performance, GoogleNet displayed slightly lower precision and recall due to misclassifications. Fine-tuning the models, exploring transfer learning techniques, or considering ensemble methods could potentially enhance their performance, particularly in mitigating misclassifications in challenging classes. In conclusion, the analysis of the four models provided valuable insights into their performance for multi-class classification. Each model showcased distinct characteristics, with varying levels of accuracy, precision, recall, and F1 score. By understanding the strengths and weaknesses of these models, researchers and practitioners can make informed decisions regarding their suitability for specific classification tasks and identify areas for further improvement. For future work, create DL and image recognition AI systems for precise trash sorting with many datasets with multi-classes and types. Investigate robotics and AI-enabled sensors for automated trash sorting. To get insights and to optimize waste management, it is necessary to combine systems powered by AI with data analytics platforms. Also, encourage teamwork, deal with privacy issues, and create intuitive mobile apps for AI-assisted trash sorting.