Dense extreme inception network-based edge detection with deep reinforcement learning for object localization in an...

The underwater environment is characterized by the vast expanses of water bodies such as seas, oceans, rivers, and lakes, together with their related ecosystems, that exist under the Earth's surface. This environment refers to unique physical properties such as pressure, buoyancy, light attenuation, and temperature, which bring about considerable adversity for observation and exploration. Object detection (OD) in the underwater environment includes the recognition and localization of different objects or entities submerged in the water. This object contains natural features like geological formations, marine life, and coral reefs, along with human-made artefacts such as debris, shipwrecks, and underwater infrastructure. Typically, object recognition in underwater environments relies on imaging technologies, namely optical cameras, SONAR, and LIDAR, which effectively operate in the complicated conditions of water. Edge detection algorithm, extracts crucial features from the surrounding aquatic landscape, fine-tuned to the unique challenges of underwater imagery, enhancing situational awareness and guiding navigation. Various techniques and algorithms are employed to detect objects in underwater imagery, as well as traditional image processing approaches, including deep learning (DL) and machine learning (ML) methods. These techniques analyze the data captured to detect distinct patterns and features related to the objects, enabling automated recognition and classification. Therefore, this study presents a new Dense Extreme Inception Network-based Edge Detection with Deep Reinforcement Learning for Object Localisation (DEINED-DRLOL) technique in an underwater environment. The primary focus of the DEINED-DRLOL technique is on the effective detection of edges and the classification of objects in the underwater environment. The Dense Extreme Inception Network for Edge Detection (DexiNed) method is employed to predict an edge map with a similar resolution. For OD, the DEINED-DRLOL technique employs the YOLOv5 method. Finally, the Q-Reinforcement Learning (QRL) method is implemented for classification. A wide range of experimentation with the DEINED-DRLOL approach is performed under the underwater OD dataset. The comparison study of the DEINED-DRLOL approach highlighted a superior accuracy value of 92.67% over existing models.