Underwater Image Processing in detection of Polymetallic Nodules

Polymetallic nodule detection underwater is faced with the extreme conditions like low visibility, various light intensity and complex textures of underwater seabed. In this study, we suggest a end-to-end machine learning and deep learning framework to enhance the underwater image analysis for polymetallic nodule detection. The mechanism of the proposed method consists of the following three stages: image recognition, boost and object detection. First, an SVM classifier is used to classify images to four types uniform illuminated, low illuminated, disturbed and ground nodules to enable a targeted processing. The low-lit images are enhanced by a CNN which can significantly outperform the traditional Contrast Limited Adaptive Histogram Equalization (CLAHE) in terms of retaining the structure of the images and enhancing visibility. Finally, a YOLOv5-based object detection model is trained with a customized dataset to effectively detect and localize the polymetallic nodules in multiple underwater situations. A comparison between the automated Nodule Hunter (with a support vector machine (SVM) classifier), CNN network, and the YOLOv5 is carried out in experiments, and the superiority results in the classification accuracy of SVM, the edge detail of the clear image of CNN, and nodule detection precision of the YOLOv5 are compared.