Underwater trash detection is vital for marine research, environmental monitoring, and underwater exploration, as trash poses significant threats to marine ecosystems and biodiversity. However, challenges such as low visibility, color distortion, and noise in underwater images hinder accurate detection. This study evaluates pre-processing techniques, including conventional methods, convolutional neural networks (CNNs), and transformer-based approaches, to address these challenges. Conventional methods, such as dehazing and histogram equalization, focus on image quality improvement, while CNN-based methods leverage deep learning to learn feature representations, and transformer-based models address complex underwater distortions by capturing long-range dependencies. Our findings indicate that conventional methods outperform both CNN- and transformer-based approaches in object detection tasks on the Trash-ICRA19 dataset. This suggests that tailored pre-processing techniques, optimized for the unique characteristics of the dataset, are essential to achieve high detection accuracy. This study emphasizes the need for dataset-specific pre-processing strategies to balance image enhancement and feature preservation, offering critical insights into improving underwater trash detection accuracy.