Enhanced recyclable waste detection and classification using artificial intelligence

Mismanagement of solid and recyclable waste is a serious problem of the 21st century. Although manual classification is necessary for waste protection and human health, it is often inefficient in terms of cost, time, and accuracy, especially in the real environment of industry, where waste inputs are not as clean as in training sets. Designed as a hybrid solution, EcoSort AI employs a two-stage approach: YOLOv8 for object detection, and ResNet18 for classification of five material categories of recyclables, with both static image and live camera detection modes supported. The dataset collection process was conducted in 6 stages of increasing complexity to build the detection model, preprocess the images to improve contrast and remove noise before inference, and create a web-based application to facilitate the upload of images, live camera recording, and reporting of detection results. The detection model achieved a mean average precision of 91.1 percent and the classification model achieved a validation accuracy of 93.0 percent. All test cases were passed in functional testing and usability testing of seven respondents gave average score of 4.17 out of 5.00, which shows a positive attitude towards the system. A correction mechanism was developed and validated at the functional level where corrections were recorded for future retraining but was not yet thoroughly evaluated with respect to the accuracy of the model, using data not included in the validation. The study concludes that the proposed hybrid architecture meets the identified research problem of automated and field-friendly waste detection and classification and recommends further expanding training data under various handheld camera conditions, further testing on held-out data to validate retraining and testing lightweight classifier architectures for real-time detection.