Improving Waste Classification Using Convolutional Neural Networks: An Application of Machine Learning for Effective Environmental Management

Waste poses significant impacts on daily human life and the environment [1]. Effective waste management is essential for addressing this issue, with waste sorting being a crucial component of the process [2]. Waste is typically categorized into inorganic, organic, and hazardous waste (B3) [3]. Upon sorting, each waste type undergoes appropriate processing: inorganic waste is recycled, organic waste is transformed into compost, and hazardous content within B3 waste is neutralized [3, 4]. Household waste, which primarily consists of organic and inorganic materials, is a common contributor to daily waste generation [5, 6].

The waste sorting process remains a challenge due to its time-consuming nature and the continuous growth of waste [7]. A prior interview conducted at a local waste bank revealed the need for a machine capable of sorting waste based on its type. One of the main hurdles faced by human sorters is the reluctance to endure prolonged exposure to the odor of waste. In contrast, machines can operate without concerns regarding the smell of garbage, making them ideal candidates for facilitating efficient waste sorting.

The development of a machine capable of classifying waste according to its type is essential for improving waste management efficiency. By employing machine learning techniques, such as convolutional neural networks (CNNs), designing an automated system that can accurately and efficiently differentiate between various waste types may be possible. This would streamline waste sorting processes and potentially mitigate the environmental impacts of waste accumulation.

Classifying waste images requires an appropriate classification method [8]. One of the classification methods that can be used is Convolutional Neural Network (CNN) [9]. CNN is a classification method in machine learning that uses many artificial neural networks and is commonly used for supervised machine learning [10]. CNN is commonly used to identify data in the form of images. The usual methods for CNN are convolution, pooling, dense, flatten, activation, and dropout [11]. These methods are often found in other classification methods, but what makes the CNN classification method unique are the convolution and pooling methods [12].

CNN implementation in machine learning requires a system that already has trained data based on the dataset [13]. The data was obtained from a dataset sourced from the Kaggle website [14]. Datasets from Kaggle can be preprocessed and simplified so that the machines can read the data [14].

The result of the implementation of CNN in machine learning is a machine that can classify images based on a predetermined type [15]. The reliability and effectiveness of the CNN classification method can be analyzed based on the accuracy and loss metric that occurs when classifying waste images [16].

Many previous studies have discussed machine learning for waste sorting using the CNN algorithm model. One of the studies is research conducted and published on the Kaggle website [17]. The research title is "Waste_classification CNN model," conducted by Bagchi [17]. The difference between Bagchi's and our research is the preprocessing and the CNN model used. The next difference is that Bagchi's research produces an accuracy of 87.10% and is still experiencing overfitting. In comparison, our study gets an accuracy of 98.92% without any overfitting seen in the analysis graph with data validation.

The following previous study is entitled Intelligent solid waste classification using deep convolutional neural networks, written by Altikat et al. [18]. This study describes the use and implementation of CNN to the problem of waste classification [18]. The difference between Altikat's and our research is the preprocessing and the complexity of the neural network. A comparison of the accuracy results in Altikat's research only reached 70% when using a five-layer neural network. The accuracy results obtained in our study was 98.92% which will be explained based on Table 1, Table 2 and Table 3.

Fahmi and Lubis [19] did a study on waste image processing in 2022, using the CNN algorithm to classify different forms of waste. The study had an accuracy rate ranging from 60% to 99%. With a 99% accuracy rate, bottles were the most accurately classified trash type utilizing CNN, followed by grass, which produced no real mistakes in the restricted system tested in this study. These findings point to the potential of CNN algorithm-based trash classification and management systems.

Shi et al. [20] investigated trash image processing in 2021, employing the CNN technique with Multilayer Hybrid feature extraction. According to the findings of this study, trash classification accuracy can reach up to 92.6%. These results indicate the efficacy of employing the CNN method in conjunction with Multilayer Hybrid feature extraction for waste classification using image processing.

Bobulski and Kubanek [21] published a study on waste image processing in 2019, leveraging the CNN algorithm to create a computer capable of capturing and detecting various sorts of rubbish. To achieve accurate trash classification through image processing, the researchers used both the AlexNet analytical model and their own analytical model. The results indicate the promise of CNN algorithm-based systems for garbage categorization and management, with 97% accuracy.

Based on many previous studies, the CNN model can be very effective for machine learning, especially for waste sorting cases. However, many only use RGB with various resizes in the preprocessing stage and still use complex CNN models.

The question from previous research is what makes the training data have the most accuracy and the most efficient classification model for handling datasets in the form of garbage images. The hypothesis believes that using complicated preparation procedures for training data can result in overfitting, primarily enhancing performance on the training data but potentially impeding generalization to new datasets. Furthermore, when image preprocessing exceeds the average level of complexity, it introduces extra distinguishing traits between objects that are beyond machine comprehension, hindering successful learning. In terms of classification models, the hypothesis posits that as the model's complexity increases, so does the time required for processing and training, potentially affecting classification performance.

The conclusion from the hypothesis is that each dataset has unique preprocessing and model complexity to achieve the desired efficiency level. Therefore, based on previous research, this study aims to determine the optimal image preprocessing techniques and CNN model complexity for effective waste classification.

The model evaluation phase produced informative data and outcomes that provide a full picture of the trained CNN model's performance and effectiveness. The model displayed amazing accuracy and efficiency in classifying photos into two categories: organic waste (O) and inorganic rubbish (R) after intensive testing and validation. The evaluation procedure included examining key parameters including accuracy and loss, which allowed for a complete examination of the model's predictive capabilities. Furthermore, the model underwent extensive validation to confirm its resilience and generalizability to new and previously unexplored data. The results reported in this section shed light on the model's ability to extract crucial features from preprocessed photos and accurately classify waste objects, therefore contributing to waste management and environmental sustainability.

Model evaluation metrics include "accuracy" and "binary cross-entropy loss." In classification tasks, accuracy is a typical evaluation parameter. The proportion of correctly categorized samples in the test set is calculated.

Accuracy in waste categorization refers to the percentage of waste items accurately classified by the model. Higher accuracy ratings suggest that the model is more accurate and trustworthy.

Binary cross-entropy is a frequent loss function for binary classification issues. It measures the difference between the anticipated probability distribution and the actual binary labels. The smaller the loss value, the more accurate the predicted probabilities are. Reduced binary cross-entropy loss allows the model to learn to produce more accurate predictions.

These metrics provide useful information about the waste classification model's performance. Accuracy aids in evaluating overall classification performance, whereas binary cross-entropy loss measures the model's prediction accuracy. You can assess the model's performance and dependability in correctly categorizing trash items by monitoring these metrics during training and analyzing them on the test set.

It should be noted that these metrics are typically utilized in binary classification jobs with two unique classes (in this case, waste or non-waste). Different assessment metrics, such as categorical cross-entropy or F1 score, may be employed if the classification problem comprises numerous classes.

In image processing, the THRESH_BINARY approach is a typical thresholding method. It divides a picture into two categories based on a threshold value. Pixels with intensity values less than the threshold are assigned the lowest value (0), whereas pixels with intensity values greater than or equal to the threshold are assigned the maximum value (255 in this case). This binary separation is ideal for jobs that require a clear differentiation, such as garbage classification.

The value 127 was chosen on the idea that it efficiently differentiates trash and non-waste regions in grayscale images. This particular value could have been derived via experimentation or domain knowledge. Finding a suitable threshold that best distinguishes between the two classes is critical for correct categorization.

In the thresholded image, the maximum value of 255 is chosen as the upper limit for pixel intensities. This value denotes the highest level of intensity, which is often used to highlight or emphasize the areas of interest. Setting the non-waste regions to the greatest value in this example aids in visually distinguishing them from the garbage regions in the final thresholded photos.

The research seeks to successfully differentiate waste and non-waste regions in photos by employing the THRESH_BINARY approach with a threshold value of 127 and a maximum value of 255. This thresholding method simplifies the image representation and focuses on the fundamental features that identify trash objects, making the subsequent classification work easier. It is crucial to note, however, that the choice of thresholding techniques and values may vary depending on the unique dataset and desired classification output, and additional research and optimization may be required to determine the most appropriate thresholding parameters.

The following are the results of the CNN schema execution based on the parameters listed in Table 1 with epochs are 30, with a dataset consisting of 25,077 waste images with the ratio of data training and data validation being 85:15.

Figure 4 and Figure 5 show two graphs of training and validation data performance analysis. Figure 4 shows a graph of machine accuracy for data classification; accuracy between training and data validation is less than 5% in each epoch. It shows that the training data and data validation are average because every time the training data has been improved, the data validation has also improved. Suppose the difference between the training and validation data is too significant at some point based on the accuracy results because the validation data does not improve while the training data is still improving; the model is overfitting [35].

4.png

Figure 4. Accuracy result

5.png

Figure 5. Loss result

Overfitting is when performance is excellent on training data, but other data is still lacking [36]. Overfitting occurs because the accuracy of the training data needs to be more precise in classifying but cannot classify waste images in general [37]. It can be found in Bagchi's research [17].

6.png

Figure 6. Accuracy and lost result of Bagchi's research

7.png

Figure 7. Range of overfitting

Figure 7 shows that the average amount of training data accuracy reduced by the average number of validation data produces a performance difference where overfitting occurs at a particular percentage difference Eq. (2). Overfitting starts at epoch 5, and maximum overfitting happens at epoch 25. Overfitting only happens when validation data does not improve while the training data is still improving.

$\frac{\sum(X \mathrm{ti}-X \mathrm{di})}{(\mathrm{En}-\mathrm{Ea})}=\mathrm{O}$            (2)

where, $X_{\mathrm{ti}}$ is the accuracy of training data on the i^th epoch, $X_{\mathrm{di}}$ is the accuracy of validation data on the i^th epoch, E_n is the number of epochs that happen, and E_a is the first epoch where overfitting occurs. The result is O, where O is the percentage value of overfitting based on the average difference between the accuracy of training and validation data. Eq. (2) only applies if the validation data does not improve while the training data is still improving.

There are many analyses about overfitting, but none of the formulas explains overfitting. The only formula known now is when to stop the learning process before becoming overfitting based on the speed of the learning rate, called early stopping Eq. (3) and Eq. (4). The learning rate will become slower and stops when the value of training data keeps improving while other data does not improve.

$\frac{\partial C}{\partial Wj}=\frac{1}{n}\sum\limits_{X}{X}j(\sigma (z)-y)$               (3)

$\frac{\partial C}{\partial b}=\frac{1}{n}\sum\limits_{X}{(\sigma (z)-y)}$               (4)

$\frac{\partial C}{\partial W j}$ and $\frac{\partial C}{\partial b}$ are two partial derivatives of the cost function, determining the neuron's learning rate [36]. Wj is the value of the jth weight, C is the cost, b is the bias, Xj is the jth input, and y is the output [36]. From this formula, only at what point to stop the learning process is known, but we cannot determine what variable makes the dataset overfitting. Based on Figure 6, it starts overfitting at epoch five and will stop around epoch five if implementing the early-stopping formula in the syntax. The following figure shows the result of early stopping based on Figure 6 cases.

The percentage of a graph of overfitting cases has yet to be determined in the general formulation. Likewise, any variables that make the dataset experience overfitting. So far, many researchers still rely on analysis from graphs to find out, so the general determination of how large the percentage difference can be classified as overfitting still needs to be discovered. The percentage difference between training and validation data can be a constant that must be examined together with what constants are appropriate for average and overfitting data conditions. The second possibility is that variables can affect changes in the average difference between training and validation data, so this variable must be determined and formulated to form a formula for determining whether the data is average or experiencing overfitting.

Overfitting happens when a model grows overly complicated and begins to memorize the training data, resulting in poor generalization of previously unseen data. The following is a more in-depth overview of how overfitting was diagnosed, overfitting signs, and the implementation of early stopping.

Overfitting can be detected by tracking training and validation loss and accuracy during the model's training process. Overfitting is usually indicated by a considerable discrepancy in performance between the training and validation sets. If the training loss continues to fall while the validation loss begins to rise or plateau, this indicates that the model is overfitting and not generalizing well to new data.

When the training loss continues to drop while the validation loss remains reasonably high or begins to grow, this is an indication of overfitting. Another indicator is a considerable gap in training and validation accuracy, where the model achieves near-perfect accuracy on the training set but struggles on the validation set.

Early stopping was used as a regularization approach to combat overfitting. It prevents the model from over-optimizing the training data by terminating the training process when the model's performance on the validation set begins to worsen. This study established a validation loss or accuracy threshold at which training is halted to avoid overfitting.

Early stopping aids in determining the ideal point during training at which the model achieves good performance on both the training and validation sets, resulting in greater generalization to unknown data.

The authors intended to achieve a balance between model complexity and generalization by diagnosing overfitting, detecting its signs, and adopting early stopping. Early stopping prevents overfitting and enhances the model's capacity to perform effectively on unseen data. Figure 6. shows visualizations exhibiting overfitting that provides further evidence and insights into the model's behavior during training.

The theory that determines what variable affects the model to be overfitting or can improve the accuracy can be obtained from comparing the parameters between the previous study, which is Bagchi's study, and our study. Two main differences occur in the preprocessing parameters, namely the resize section and the color scale of the training data. In our study, the resize used was 50 pixels, while the previous study used 200 pixels. Then the color scale used in our study is grayscale, while the previous research is RGB. It shows that the input shape used is different; in our study, the input shape is 50x50x1, 1 indicates the color scale used is grayscale based on Eq. (1) while the previous research has 200x200x3, 3 indicates the color scale used is RGB based on Eq. (1).

Most previous journals claimed that resizing could affect the performance of the CNN model [38, 39]. It was also explained that the most effective resizing is a size close to the original image size [38]. So, there is no correlation. If the size is bigger, then the performance of the CNN model will be even more outstanding. There is also a theory that the larger the size value, the greater the accuracy and the more time used, especially when using the CNN model [39]. However, our study contradicts the two opinions because in our study, the average size of the original image is above 200x200 pixels, so it is not fulfilled for the first and second opinions because research has better accuracy using 50x50 pixels rather than 200x200 pixels. Resizing will have an effect, but the indicator that makes the performance better is the suitability of the size of the data image for a particular dataset, different datasets used will also have different efficient image sizes.

8.png

Figure 8. Example of the graph when implementing early stopping

Most previous journals also claimed that grayscale would work better [40]. However, many studies still use RGB [18, 41, 42]. The case study using the same dataset as the study [17] shows a difference in color scale. In research [17], they use RGB so that the final variable input shape entered into the CNN model is 3, while in our study, this research use grayscale so that the final variable input shape entered into the CNN model is 1. Grayscale has better performance and is also more efficient in its use [40]. The accuracy in classifying by the machine will increase if the object has a transparent edge and contrasts color with the background. The convolutional can give weight and bias value with a smaller range due to the limited color available while the contrast is still maintained. However, this opinion still needs improvement when comparing different datasets and models. The use of grayscale can be more effective if using specific datasets and RGB if done on other datasets [18, 40-45 ]. The following is Figure 8 explaining early-stopping when implemented into machine learning. It shows that early stopping only stops machine learning when overfitting might happen.

The following comparison is the CNN model based on the neural network's architecture.

Table 3. CNN scheme of Bagchi's research

A comparison of the CNN model between the three studies (Table 1, Table 2, and Table 3.) can be seen from the number of layers built. The more layers used, the more complex the machine classifies objects. The level of model complexity can affect machine learning performance if the existing dataset is limited [46]. Several studies explain that the more complex a model is, the more vulnerable it will be to reduced accuracy performance [46]. However, some argue that complex models can improve machine learning performance because the function of complexity can solve problems that occur [47]. For example, CNN combined with LSTM will be good because the function of the two models, such as CNN as a model that functions as a recognition skill, LSTM as a model that provides time series data, and the combination serve as a calculation speed in problem-solving [44]. This study (Table 1), compared to the study (Table 2), had better accuracy results even though the depth and complexity of the neural network were more straightforward in this study. The complexity and depth of the CNN model architecture have an influence, but each dataset has a unique value, so it cannot be applied to different datasets.

After the result of training data compare with validation data is no resulting overfitting, the next step is to see the accuracy result tested using test data. The acquired result is an accuracy of 98.92% and loss of 4,03%. Figure 9 is the result of evaluating machine learning using test data.

The three CNN models mentioned in Table 1, Table 2, and Table 3, prior research by Fahmi and Lubis [19], Shi et al. [20], and Bobulski and Kubanek [21], we can look at model architecture, hyperparameters, performance metrics, and overfitting. The three CNN models in the table differ in terms of model design, including the amount of Conv2D and Dense layers, their sizes, and the usage of additional layers such as MaxPooling2D and Flatten.

Fahmi and Lubis [19] design produced an accuracy rate ranging from 60% to 99%. Shi et al. [20] used the CNN technique with Multilayer Hybrid feature extraction, whereas Bobulski and Kubanek [21] used the AlexNet analytical model as well as their own.

9.png

Figure 9. Result with test data

Each CNN model in Table 1, Table 2, and Table 3 and earlier research had distinct hyperparameter values, including kernel sizes, input shapes, activation functions, and optimization strategies, based on hyperparameters.

According to Table 1, Table 2, and Table 3 performance metrics, the CNN models achieved binary cross-entropy loss and employed accuracy as a performance parameter. Fahmi and Lubis [19] observed trash classification accuracy rates ranging from 60% to 99% using their CNN algorithm. Shi et al. [20] used the CNN method with Multilayer Hybrid feature extraction to achieve waste categorization accuracy of up to 92.6%. Bobulski and Kubanek [21] specific performance measurements have up to 97% accuracy. This study assumes that only Bagchi's work is overfitting because it appears very clearly in Figure 6, while other work doesn’t show any evidence.

Differences in the results obtained from the models and studies can be attributed to various factors, including variations in model architectures, hyperparameter values, dataset characteristics, and data preprocessing techniques. Additionally, variations in the performance metrics used and the specific waste classification tasks tackled by each study could contribute to the differences observed.

However, when compared to the machine learning time required to classify, it is calculated that this research could be faster. One of the most significant factors that causes time to slow down is the ability of the computer to execute syntax commands. In this study, the computers used had specifications incapable of carrying out heavy executions, especially on the display device. Here are the specifications of the computer used. System Manufacturer: LENOVO, System Model: 10HV002JIA, BIOS: LENOVO BIOS Rev: M0KKT17A 0.0 (type: BIOS), Processor: Intel(R) Core(TM) i3-4170 CPU @ 3.70GHz (4 CPUs), ~3.7GHz, Memory: 4096MB RAM, Available OS Memory: 4006MB RAM, Page File: 3923MB used, 1553MB available, and the display device specification is ard name: Intel(R) HD Graphics 4400, Manufacturer: Intel Corporation, Chip type: Intel(R) HD Graphics Family, DAC type: Internal, Device Type: Full Device (POST), Device Problem Code: No Problem, Driver Problem Code: Unknown, Display Memory: 2115 MB, Dedicated Memory: 112 MB, Shared Memory: 2002 MB, Current Mode: 1440 x 900 (32bit) (60Hz), HDR Support: Not Supported, Display Topology: Internal, Display Color Space: DXGI_COLOR_SPACE_RGB_FULL_G22_NONE_P709, Color Primaries: Red(0.653320, 0.336914), Green(0.322266, 0.610352), Blue(0.151367, 0.064453), White Point(0.313477, 0.329102), Display Luminance: Min Luminance=0.5000000, Max Luminance=270.000000, MaxFullFrameLuminance=270.000 000.

The value is inversely proportional to accuracy compared with the loss percentage, like in Figure 5, because the validation data is greater than the training data. This incident proves that a good result of any classification method is when the accuracy percentage has a high value. In contrast, the loss percentage has a small value while training and validation data keep improving every epoch.

The test data results also show that our research's accuracy is better than previous studies using preprocessing and the previously described model with a value of 98.92%. It also shows that the garbage dataset [14] is more efficient using the same preprocessing parameters and CNN model as our study. For the syntax, it uses references from previous studies [17] as control variables but has different preprocessing values and the complexity and depth of the model.