Classification of recyclable wastes with deep learning models: A comparison on the effect of dataset size

Recycling plays a crucial role in promoting environmental sustainability by reducing the demand for natural resources, minimizing waste accumulation, and limiting the emission of pollutants into the environment. As global populations increase and consumption habits intensify, the management of waste has become a critical challenge for governments, industries, and communities worldwide. One of the most essential steps in effective waste management is the proper classification of recyclable materials. This step ensures that waste streams are sorted efficiently, facilitating easier recycling and preventing reusable materials from being lost in landfills. Traditionally, waste classification has relied on manual processes that are time-consuming, error-prone, labor-intensive, and expensive. As such, there is a growing interest in automating waste classification processes, and recent advances in deep learning have provided promising approaches to address this need. This thesis explores the use of deep learning models for the classification of recyclable waste materials based on visual input. Convolutional Neural Networks (CNNs), which are known for their strength in image classification tasks, are employed and evaluated for their performance in distinguishing between different types of recyclable waste. The research examines how dataset size, class balance, and model architecture affect classification performance. Two publicly available datasets—Garbage Classification and Garbage-Dataset—were utilized in order to test and compare the performance of different CNN architectures under varying data conditions. The Garbage Classification dataset includes a total of 2,890 images divided into six classes: cardboard, cam, metal, paper, plastic, and trash. This dataset is characterized by significant class imbalance, with some categories containing many more samples than others. Such imbalance can negatively affect model training, as the network may become biased toward more dominant classes. To address this issue, data augmentation techniques were employed, including random rotation, horizontal flipping, and image scaling. These augmentations were applied to artificially expand the dataset and enhance its diversity, helping the model generalize better during training. Additionally, image resizing and normalization were applied to prepare the images for input into the CNN models. The second dataset, Garbage-Dataset, contains 9.660 images and offers a more balanced class distribution across the same categories. With more samples and greater variability, this dataset serves as a benchmark for evaluating model generalization and robustness under more realistic conditions. It includes various background environments and object perspectives, making the classification task more challenging but also more representative of real-world scenarios. This dataset allows for a deeper analysis of model performance and helps to uncover the strengths and weaknesses of different network architectures. In this research, four state-of-the-art CNN architectures were selected for evaluation: VGG16, InceptionV3, MobileNetV2, and DenseNet121. These models were chosen due to their widespread usage, architectural differences, and proven performance in image classification benchmarks such as ImageNet. To accelerate training and improve model accuracy, transfer learning wasapplied. The models were initialized with pre-trained weights from ImageNet and fine-tuned using the recyclable waste datasets. This approach allows the models to retain general visual features learned from large-scale image datasets while adapting to the specific characteristics of waste images. To ensure robust evaluation, a five-fold cross-validation strategy was implemented for each model and dataset. Classification performance was measured using accuracy as the primary metric. However, to provide a more comprehensive evaluation, additional performance indicators such as precision, recall, and F1-score were calculated. These metrics were particularly important for understanding model performance on underrepresented classes in the imbalanced dataset. Confusion matrices and class-wise performance metrics were also used to highlight common misclassifications and analyze the causes behind them. The experimental results show clear differences in model performance depending on the dataset size and structure. On the Garbage Classification dataset, which is smaller and more imbalanced, the VGG16 model achieved the best accuracy of 85.05%, outperforming the other architectures. DenseNet121, MobileNetV2, and InceptionV3 followed with accuracies of 82.63%, 82.49%, and 80.42%, respectively. This suggests that simpler models such as VGG16 may be more suitable for smaller datasets due to their lower risk of overfitting. In contrast, on the larger Garbage-Dataset, the performance rankings changed. The DenseNet121 model led with an accuracy of 90.75%, followed by MobileNetV2 at 89.67%, InceptionV3 at 88.41%, and VGG16 at 86.09%. These results indicate that deeper and more complex models benefit from the availability of more data, which allows them to utilize their larger capacity to learn intricate patterns and features. It also confirms the importance of choosing a model architecture that matches the scale and characteristics of the available training data. Beyond raw accuracy, performance on individual classes revealed interesting insights. For example, certain classes such as cardboard and glass were consistently well-identified across all models, while metal and trash were more prone to misclassification. These errors can often be attributed to the visual similarity between certain waste types or the presence of background clutter in the images. Further error analysis showed that class confusion often occurred between plastic and metal, and between paper and cardboard, likely due to overlapping textures and colors. To improve model interpretability, techniques such as Grad-CAM (Gradient-weighted Class Activation Mapping) were explored. These techniques help visualize which parts of an image the model focuses on during classification, offering transparency and aiding in debugging model decisions. Such explainability is critical in practical applications where users need to trust and understand model outputs. The practical deployment of these models was also considered. In real-world applications, especially within smart waste management systems, models must be not only accurate but also efficient in terms of memory usage, processing time, and energy consumption. Among the evaluated models, MobileNetV2 stood out as the most suitable for embedded systems and mobile applications due to its lightweight architecture and competitive performance. In terms of environmental policy and societal impact, automated waste classification can support broader sustainability initiatives by enabling municipalities to manage recyclable waste streams more effectively. It can also reduce reliance on manual labor, thus lowering costs and health risks associated with manual sorting processes. These systems may further empower communities by increasing awareness around recycling practices and contributing to cleaner urban environments. Limitations of the study include the scope of the datasets and the reliance on visual data alone. In real-life scenarios, other data types such as material composition or sensor readings could complement visual classification for improved accuracy.Additionally, feature studies could explore the use of ensemble learning, semi-supervised training, or generative approaches such as GANs (Generative Adversarial Networks) to expand datasets and improve model robustness. Moreover, model explainability will become increasingly important as these systems are adopted more widely. For this reason, researchers must continue to investigate methods for making CNN decisions interpretable to users. Collaborative projects that bring together data scientists, environmental engineers, and policymakers will be essential in translating technical advancements into impactful, real-world waste management solutions. In conclusion, this study demonstrates the effectiveness of CNN-based deep learning models in classifying recyclable waste materials. By evaluating different architectures across datasets of varying size and quality, the research provides valuable insights into how model choice and data characteristics influence classification success. The findings support the use of deep learning for automating waste sorting processes and contribute to the broader goals of environmental sustainability and smart city development. Feature research should continue to refine these models, develop more diverse and realistic datasets, and pursue deployment strategies that balance accuracy with computational efficiency. The integration of such intelligent systems into waste management infrastructure could dramatically improve recycling rates, reduce operational costs, and support global efforts in resource conservation. In addition to evaluating individual CNN architectures, this study proposed and implemented a novel hybrid model based on feature fusion, referred to as the Feature Fusion Hybrid Model. The hybrid approach aims to leverage the complementary strengths of multiple pre-trained CNNs—namely VGG16, InceptionV3, DenseNet121, and MobileNetV2—by extracting high-level features from each model and concatenating them into a unified feature representation. This combined feature vector was then passed through fully connected layers for final classification. The hybrid model was trained using five-fold cross-validation on both the Garbage Classification and Garbage-Dataset. Experimental results showed that the hybrid model achieved a test accuracy of 85.88% on the Garbage Classification dataset, surpassing all individual base models. On the more comprehensive Garbage-Dataset, the hybrid model reached an accuracy of 91.26%, exceeding the best-performing individual model (DenseNet121 with 90.75%). These results demonstrate the potential of feature fusion in improving classification performance by integrating diverse feature representations from multiple CNN architectures, leading to more robust and accurate waste classification systems.