Introduction: Plastic waste pollution is one of the major environmental challenges, and the automatic detection and localization of such waste in images can provide the data required for waste monitoring, cleanup planning, and targeted and intelligent waste management. The aim of the present study was to develop and evaluate an object detection model based on the new YOLO26n architecture, introduced in 2026, using a transfer learning approach to simultaneously detect and localize five important categories of plastic waste, including plastic bottles, plastic cups, plastic packets, plastic straws, and polythene, in images with real-world environments and backgrounds. The model was trained on the new PlasticVision-BD dataset, for which no article reporting modeling on the dataset was identified at the time of writing.
Materials and Methods: After preprocessing, the PlasticVision-BD dataset comprised 15,098 images and 18,862 plastic objects distributed among three groups: training, validation, and test. Prior to model development, the independence of the three dataset groups was assessed by examining exact image duplication using SHA-256 digital fingerprints and perceptual similarity using pHash. No exact duplication or extensive overlap that could compromise the independence of the groups was observed; this assessment increased confidence in the model evaluation results. The model was fine-tuned using pretrained YOLO26n weights through a transfer learning approach. Training was stopped after 36 epochs following the observation of relative performance saturation during the final epochs, and both the model with the best weight obtained during training and the model representing the final training state were saved. Both models were evaluated on test-set images that had not been seen by the model during training using Precision, Recall, F1 score, mAP50, and mAP50-95. In a separate evaluation, model processing speed was assessed by applying the model to all images in the dataset.
Results: The model with the best weight outperformed the model with the final training state across all evaluated metrics. On the test set, the best model achieved a Precision of 89.53%, a Recall of 76.88%, an F1 score of 82.72%, an mAP50 of 80.80%, and an mAP50-95 of 57.39%. Among the five categories, the model showed the best overall performance in detecting and localizing plastic bottles, whereas polythene was the most challenging category to detect. Visual evaluation on the test images also demonstrated the model's ability to simultaneously identify the type and location of one or more plastic waste objects within a single scene. In a separate evaluation of processing speed, the model processed all images in the dataset in approximately 25 minutes, whereas model training took nearly 69 hours.
Discussion: Comparison with previous studies indicated that, with respect to the key mAP50 metric, the model developed in the present study showed good performance standards; however, such comparisons should be interpreted in light of differences in datasets, number of categories, imaging conditions, and training settings. In addition, the relatively high processing speed, the capability for repeated use of the model after training, and its practical applicability in multiple systems, despite the relatively considerable time cost of the training stage, fully justify model training. The results indicate that the model can be used as an intelligent component in environmental image-monitoring systems, waste sorting and recycling lines, cleanup operation prioritization, automated and semi-automated collection and robotic systems, and mobile phone-based software, including decision-support systems for source-separated waste sorting. However, to move toward practical deployment at a real-world scale, future studies should test the model on diverse field images, under different lighting and background conditions, and using data from operational environments, and should investigate and improve the model's generalization capability to conditions outside the training dataset.