Waste Detection in underwater using deep Vision Transformers

The detection of waste in underwater environments is a critical task for marine conservation, yet it poses significant challenges due to the complex and dynamic nature of these settings. Factors such as low visibility, varying light conditions, and diverse waste materials make traditional detection methods, like Convolutional Neural Networks (CNNs), less effective. To address these challenges, this research explores the application of Deep Vision Transformers (DVTs) for underwater waste detection. DVTs are particularly well-suited for this task due to their advanced features, including the ability to capture global contextual information and model long-range dependencies within an image. Unlike CNNs, which focus on local features, DVTs utilize self-attention mechanisms to effectively process and prioritize different parts of an image. This leads to robust feature extraction, which is crucial for accurately identifying waste in complex underwater environments. The Project employs techniques such as transfer learning and data augmentation to enhance the performance of DVTs on the limited datasets available for underwater imagery. Fine-tuning the model with labeled underwater images further improves detection accuracy.