Improving Waste Sorting Accuracy with Deep Learning and VGG-16 Architecture
Shiva Mehta, Anubhav Bhalla · 2025
Waste management as an essential component of sustainability is the foundation for the implementation of an intelligent system for automated waste classification as the key link in the processes of recycling. Therefore, to address this problem, this work presents a solution based on post-processing of images through a deep learning approach in the form of the VGG-16 network with a suitable estimator to categorize waste as biodegradable, recyclable or hazardous material. The multi-class dataset was used to train and test the model and incorporates some state of art preprocessing and augmentation methods. The results proved the model’s high accuracy and specific accuracy of each type of the waste achieving such rates 92.5% overall accuracy; 93.2% of precision for biodegradable waste; 92.5% of recall; the F1 score of 92.8%. In the same regard, recyclable waste yielded precision of 91.7 % , 90.8% recall and F1-score of 91.2 % while the hazardous waste yielded a precision of 94.5 % , 92.3% recall and F1-score of 93.4%. Receiving Operating Characteristic (ROC) showed the performance of the model in the next step with an area under the curve of 0.96, which means a better classifier separates between classes. Metrics of convergence based on the training and validation set were established to reflect an improved model with training accuracy increasing from 60% to 95 % in 20 epochs while validation accuracy increased from 58% to 92%. Furthermore, identifying the misclassified images, there were significantly few numbers of false positives and false negatives for the categories of recyclable-and hazardous wastes; 10 and 5, respectively.It should also be noted that our classifier based on the VGG-16 model showed higher accuracy (93.9%) and faster computations than the baseline models SVM (accuracy 78.5%) and ResNet-50 (accuracy 91.8%). These conclusions prove the effectiveness of the model and its adaption for using in waste management systems that are needed for automated and efficient recycling of materials. As for future work, there will be work on small tweaks of the currently identified misclassification issues and improvement of the system for IoT environments.