Research On Underwater Image Enhancement And Target Detection Algorithm Based On Deep Learning

The development of underwater drones has led to an explosive increase in the amount onto data of underwater optical images,and the research interest in underwater images is growing.Underwater optical images can provide important support for many underwater operations.For example,technicians can inspect and repair underwater artificial facilities through real-time video returned by underwater robots,biologists can conduct research on marine life through underwater image pairs,and geologists can detect resources through seabed images.In these tasks,underwater image processing algorithms and underwater target detection algorithms are the most commonly used auxiliary tools.However,the special environment under water will cause serious degradation of the image.Traditional image enhancement algorithms are difficult to apply to complex underwater images,and the existing mainstream target detection models also have many false detections and missed detections when applied to underwater images.Based on deep learning,this paper explores underwater image processing and underwater target detection algorithms.The specific research contents are as follows:First,in response to the lack of datasets for current supervised learning underw ater image enhancement algorithms,this paper proposes a GAN(Generative Adversar Ial Networks).Combined with the underwater image imaging model,the correspo nding underwater style images are generated by using the RGB-D image dataset col lected on land,which constitutes a paired underwater image enhancement dataset.Second,in view of the phenomenon that traditional image enhancement algorithms are difficult to apply to underwater images,this paper proposes a supervised learning underwater image enhancement model.Combined with WPT(Wavelet Packet Transform)and CNN(Convolutional Neural Networks),the detailed feature extraction ability and deblurring ability of the model are strengthened,and the style loss function is introduced to increase the color correction ability of the model.Thirdly,in view of the missed detection and false detection of underwater fuzzy targets,overlapping targets and small targets by popular algorithms,this paper proposes an improved algorithm based on Faster R-CNN.The FPN(Feature Pyramid Networks)structure is introduced into the backbone feature extraction network to enhance the model’s detection ability for small targets.Replace NMS(Non Maximum Suppression)with WBF(Weighted Boxes Fusion)in the RPN to enhance the model’s ability to distinguish overlapping targets.The double-head structure is used to replace the original single-head structure,and the coordinate adjustment and classification tasks of the recommendation box are performed separately to enhance the model’s ability to locate fuzzy targets.Experiments show that the underwater image generation model proposed to this paper can effectively expand the training set of the supervised underwater image enhancement model.The underwater image enhancement model proposed to this paper has advanced performance in both subjective vision and objective indicators.The underwater target detection model proposed to this paper has a higher m AP(Mean Average Precision)than Faster R-CNN and the comparison algorithm.