Identificação de lixo hospitalar por meio de imagens de smartphones com deep learning

Abstract

The inadequate management of hospital waste poses a significant risk to urban environments, which has been further exacerbated by the COVID-19 pandemic. This study has created an application utilizing deep learning techniques to identify instances of improper waste disposal. A meticulously curated database was compiled, encompassing a wide variety of images from various sources. Segmentation masks were meticulously generated for each image, and data preprocessing and augmentation techniques were Applied to create a secondary dataset. Subsequently, we trained twelve distinct models using specialized techniques to enhance their performance, employing the YOLOv8 architecture for image segmentation. The final model of choice was YOLOv8x, which yielded the most promising results in key metrics such a recall rate of 0.62, an mAP50 of 0.67, and an mAP50-95 of 0.43, particularly when trained on the augmented dataset. This selected model was then deployed on a local server and seamlessly integrated with the application through asynchronous web requests. The application itself was developed using Swift, providing users with the capability to submit images captured by their câmeras for server-based analysis. The server draws masks to identify instances of improper waste disposal and returns the modified image to the user through the application.