HARNESSING DEEP TRANSFER LEARNING WITH IMAGING TECHNOLOGY FOR UNDERWATER OBJECT DETECTION AND TRACKING IN CONSUMER ELECTRONICS
Abstract
Consumer electronics like action underwater drones and cameras commonly include object detection abilities to automatically capture underwater images and videos by tracking and focusing objects of interest. Underwater object detection (UOD) in consumer electronics revolutionizes interactions with aquatic environments. Modern consumer gadgets are increasingly equipped with sophisticated object detection capabilities, from action cameras to underwater drones, which allow users to automatically capture clear videos and images underwater by tracking and identifying objects of interest. This technology contributes to user safety by enabling devices to avoid collisions with underwater obstacles and improving underwater videography and photography quality in complex systems simulation platforms. Classical approaches need a clear feature definition that suffers from uncertainty due to differing viewpoints, occlusion, illumination, and season. This paper focuses on developing Deep Transfer Learning with Imaging Technology for Underwater Object Detection and Tracking (DTLIT-UOBT) techniques in consumer electronics. The DTLIT-UOBT technique uses deep learning and imaging technologies to detect and track underwater objects. In the DTLIT-UOBT technique, the bilateral filtering (BF) approach is primarily used to improve the quality of the underwater images. Besides, an improved neural architectural search network (NASNet) model derives feature vectors from the preprocessed images. The DTLIT-UOBT technique uses the jellyfish search fractal optimization algorithm (JSOA) for the hyperparameter tuning process. Finally, the detection and tracking of the objects can be performed by an extreme learning machine (ELM). A sequence of simulations was used to authorize the performance of the DTLIT-UOBT model by utilizing an underwater object detection dataset. The experimental validation of the DTLIT-UOBT model exhibits a superior accuracy value of 95.71% over other techniques.
