Underwater Objects Detection Based on a Multi-Stage Deep Learning Framework

The challenges of underwater object detection are derived from complex environmental conditions, including light scattering, absorption, and turbidity. The deep learning approaches have enhanced the detection of objects in these low-visual conditions. This work presents a multi-stage object-detection framework for the underwater environment that performs well on the Semantic Segmentation of Underwater Imagery (SUIM) benchmark. To begin with, there is the adaptive Multi-Scale Retinex with Color Restoration (MSRCR) algorithm, which improves image quality by correcting color distortions and increasing contrast. Second, an augmented YOLOv8 model (with a ResNet-50 backbone and the Convolutional Block Attention Module (CBAM)) is used to extract powerful features for object detection in low-light conditions. Lastly, a LightGBM classifier selects initial detections using contextual information to reduce false positives. The proposed model is evaluated on the SUIM dataset, with ground-truth segmentation masks converted to bounding boxes according to standard COCO protocols for detection-based training and evaluation. Comparative experiments against reimplemented YOLO-based underwater detectors demonstrate that the proposed model achieves a macro-average mAP@0.5 of 86.33%, outperforming YOLOv8-nano (80.13%), YOLOv7+enhancement (84.23%), and AIT-YOLOv7 (83.4%) on the SUIM benchmark under similar conditions.