Research On Underwater Target Detection Algorithm Based On Deep Learning

Underwater object detection is a challenging field of research primarily due to the unique characteristics of the underwater environment and the difficulty in acquiring data.Light attenuation in water leads to the loss of details and edge information in captured images.Additionally,the high cost of data collection limits the quality of available datasets.This paper aims to explore methods for improving underwater object detection algorithms through in-depth analysis and research.The specific contents of this paper are as follows:Firstly,addressing the difficulty in acquiring datasets for underwater object detection,a dataset augmentation algorithm called MUOD is proposed to adapt to the imaging characteristics of underwater optical images.This algorithm effectively enhances dataset diversity and alleviates the challenges of dataset acquisition and low image quality.Experimental analysis on publicly available datasets demonstrates that this algorithm significantly improves the detection accuracy of object detection algorithms.Secondly,addressing the unique imaging characteristics of underwater optical images,improvements are made to the FCOS object detection algorithm.The regular convolutions in the backbone network are replaced with deformable convolutions to enhance the feature extraction capabilities when dealing with blurred underwater optical images.The feature pyramid network and detection network are optimized through neural architecture search to improve the efficiency of utilizing features extracted by the backbone network.The CIo U Loss is employed as a new loss function to enhance coordinate regression accuracy.Through experiments on publicly available datasets,the improvement effects of each module on the FCOS algorithm are analyzed,and a comparison is made with mainstream object detection algorithms to demonstrate that this algorithm outperforms existing state-of-the-art algorithms in terms of accuracy.Finally,considering the demand for lightweight underwater object detection algorithms,a lightweight high-speed detection algorithm is designed.Mobile Net V2 is employed as the backbone network,and a novel receptive field fusion mechanism is incorporated to ensure both model lightweight and detection accuracy.A high-quality bounding box classification and regression network is designed to improve the quality of generated detection boxes.An enhanced non-maximum suppression algorithm is utilized to improve the detection rate of overlapping objects.Experimental validation on public datasets demonstrates that this algorithm achieves a slightly reduced detection accuracy while offering faster detection speed and a more lightweight model.