Underwater Scene Enhancement And Object Detection Based On Deep Networks

Underwater object detection is of great significance for ocean exploration and monitoring,autonomous underwater vehicles and other underwater applications.Influenced by the complexity of the marine environment,the actual underwater images suffer from serious degradation,mainly in terms of visual blur,color distortion,low contrast,etc.This seriously affects underwater object detection performance.It is extremely important to enhance the visual quality of underwater images as a preprocessing operation,so as to improve the accuracy of underwater object detection.With the rapid development of deep learning,object detection has brought a qualitative leap.It is of great theoretical and practical value to explore the application methods of image enhancement and object detection in underwater scenes.The major contributions are as follows:First,a fusion based underwater scene enhancement model is proposed.The model is in the form of an end-to-end deep convolutional neural network.With the characteristics of underwater image degradation taken into consideration,the structure of the gated fusion network is established through three preprocessing operations and fusion strategies.An enhanced underwater visual perception dataset(EUVP)containing reference images is utilized to conduct a comprehensive qualitative and quantitative analysis in comparison with several existing underwater scene enhancement algorithms.The effectiveness,adaptability and stability of the proposed model are verified,and the limitations of the existing underwater scene enhancement methods are summarized.Second,to meet the requirement of the underwater object detection task in terms of speed and accuracy,YOLOv5 is adopted as the core of the underwater object detection algorithm.A variety of image processing techniques are introduced and the model structure of YOLOV5 is analyzed.In view of four configurations of YOLOv5 with different network depths and network widths,qualitative and quantitative evaluations are conducted in real underwater biological datasets to demonstrate and summarize the performance differences of the four configurations.In view of the relationship between scene enhancement and object detection,an underwater scene enhancement model in the form of a fusion based deep convolutional neural network and YOLOv5 are combined to conduct experimental evaluations.The correlation between them is verified.Experimental results show that eliminating color bias and blurred vision can improve the performance of the subsequent object detection task.In summary,this thesis provides not only a comprehensive study on underwater scene enhancement and underwater object detection,but also a systematical analysis of the correlation between enhancement effect and detection performance.The aim is to develop underwater scene enhancement methods and thus improve the accuracy of object detection.This provides a new idea for further studies on underwater image processing.