Deep Learning Based Underwater Image Enhancement And Fish Recognition

China has a vast marine territory and rich and diverse marine resources,of which fishery resources are an important part.The study of fish detection and identification is of great importance to assist fishery investigation,fishing and processing.However,images acquired in complex underwater environments are usually heavily affected by various noises,which can negatively affect the execution of underwater target detection tasks.The variety of fish forms in the underwater environment makes fast and accurate localization and identification of fish targets in the underwater environment challenging in many ways.Traditional fish identification methods usually use a strategy of combining manually designed features with machine learning classifiers.However,manual features suffer from many problems such as lack of generality of features,difficulty in extraction and time consuming.In recent years,deep learning techniques have made remarkable progress in the field of image classification,and they have excellent performance in image detection and recognition.Therefore,in this thesis,we use deep learning methods for underwater image enhancement and fish detection and classification research,and the main work is as follows:1.The research and development on underwater image enhancement and underwater fish detection at home and abroad are thoroughly investigated.The basic principles of convolutional neural networks,methods of image enhancement,and processes of target detection are outlined,and the basic structures and principles of the two commonly used target detection algorithms are dissected with emphasis on these contents to provide the theoretical basis for the next underwater image enhancement and fish detection algorithms.2.For the problems of low quality of underwater images,such as low contrast,blurred details,etc.,and defects in local and detailed enhancement of underwater images,this thesis uses a model based on cycle-consistent generative adversarial networks(CycleGAN)to perform underwater The generators in this model are capable of The generator in this model can realize the transformation between two domains,and the discriminator can determine whether the input image is the image that matches the transformed domain.Compared with traditional image enhancement methods,this method can better preserve the local detail information of the image while performing color restoration and defogging.3.The problem of low accuracy of generic YOLOv5 in underwater target detection task.YOLOv5 is improved for the target detection task in this thesis.The feature extraction network of YOLOv5 is improved by incorporating the CBAM attention mechanism,which aims to improve the extraction of relevant features and suppress the irrelevant background information;since the original CIOU loss function used in YOLOv5 does not take into account the mismatch between the required real frame and the predicted frame,the predicted frame may "wander around" during the training process."This thesis will use the loss function SIOU to replace the loss function of the original YOLOv5,which takes into account the vector angle between the required regressions and redefines the penalty index,which can accelerate the convergence speed and improve the efficiency.Through experimental comparison of the improved underwater fish detection algorithm,an accuracy of 85.7%is achieved with a loss of 10 points of FPS on the dataset,which is 4.3 percentage points higher than the original YOLOV5;to improve the performance of the underwater fish detection algorithm in practical application scenarios,MobileNetv3 is used as the backbone network of YOLOv5,and the improved MobileNetv3-YOLOv5 improves the FPS from 34 to 56 with a loss of 1.3 percentage points of accuracy.