AquaVision: A Diffusion-Augmented Transformer-Driven Multi-Stage Enhancement Framework for Object Detection in Occluded Underwater Environments

Underwater object identification is becoming an important tool for maritime research, environmental monitoring, and autonomous navigation. However, light absorption, dispersion, low visibility, and significant colour distortions make it much harder to find things, which often lowers the quality of underwater images. It is still hard to get reliable, real-time object recognition since traditional augmentation methods tend to overcorrect or fail when the water conditions change. The proposed framework is a complete deep learning-based improvement and detection pipeline in this study that makes underwater photographs much clearer and easier to use. The method has three important parts: a colour correction network that uses VGG16 and attention mechanisms (CBAM) to get the right colour tones back; a resolution enhancement unit that is based on a modified U-Net architecture and improved with Swin Transformers to capture local and global image features; and a denoising component that uses Denoising Diffusion Probabilistic Models (DDPM) to get rid of artefacts while keeping structural details. After that, a YOLOv7-based detection method that works with DeepSORT is used on the better photos to make sure that the locations and movements are tracked correctly. The examination of the enhancement strategies on the TrashCan and ICRA datasets shows that they make a big difference in precision, recall, and mean average precision (mAP). This shows that they work to increase detection performance. This framework can be used for a lot of things that happen underwater, like underwater robotics, marine biodiversity assessment, search-and-rescue operations, and oceanographic research.