Type Deep Learning Model for Multi-Label Waste Classification in Canal Environments: A Comparative Study with CNN Architectures

Abstract

The escalating environmental degradation caused by waste underscores the necessity of developing intelligent and sustainable management systems. This study introduces a deep learning–based framework with proposed a modified ConvNeXt architecture enhanced by a two-layer non-linear MLP classification, specifically designed for multi-object waste classification in canal environments. Specifically, ConvNeXt-CNN is introduced as the primary backbone for extracting visual features from waste images. Then, a modified Multi-Layer Perceptron (MLP) is employed to transform these features into multi-label predictions. To optimize the model’s generalization capability in recognizing the complexity of waste images, a hybrid data augmentation technique combining SMOTE and MixUp was applied during training. The proposed approach was then compared with ten fine-tuned Convolutional Neural Network (CNN) architectures, ResNet18, ResNet50, VGG16, VGG19, DenseNet121, MobileNet_v2, and EfficientNet (B0, B1, B2, and B3), and evaluated using accuracy, precision, recall, and F1-score metrics. The experimental dataset comprises 855 waste images containing a total of 2,662 annotated objects across 18 categories, including Bamboo, Beverage Carton, Cardboard, Fabric, Glass Bottle, Inorganic Waste, Kite, Leaf, Metal, Organic Waste, Paper, Plastic, Plastic Bottle, Plastic Cup, Residual Waste, Rubber, Small E-waste, Styrofoam, and Wood. The results show that the fine-tuned ConvNeXt achieved the best performance with an F1-score of 0.99, surpassing DenseNet121 (0.95), ResNet18 (0.91), and VGG16 (0.94). The ConvNeXt model demonstrated its robust capability by achieving consistently high identification scores across majority 18 waste categories. When it came to training efficiency, the fine-tuned MobileNetV2 model proved to be the top performer, outclassing ten other pretrained models, with a training time of 13.35s per epoch.  Results exhibit that finetuned ConvNext outperforms in terms of accuracy, recall, precision, and F1-score. In conclusion, Integrating ConvNeXt and MLP for multi-object waste classification effectively supports intelligent waste management, enabling practical real-world deployment in smart bins, Material Recovery Facilities, and IoT-integrated urban waste systems.