Research On Fish Detection And Recognition In Underwater Low-light Images

With the growing attention to marine resources,applying modern marine information technology to establish a modern marine industry system has become one of the most important initiatives of our country.As the most widely distributed and representative among marine biological resources,fish has also become a major research object for scholars at home and aborad.During the process of investigation,fish detection and recognition is the fundamental steps.However,due to the limitations of the special underwater imaging environment,underwater images often suffer from low contrast,color distortion,uneven illumination and other quality degradation,resulting in loss of details and difficulties in automated analysis;in addition,the different morphologies,sizes,and various categories of fish bring serious challenges to detection and recognition.In order to overcome these difficulties,restore the quality of underwater images and achieve accurate and stable detection and recognition of fish among low-quality underwater images,the main contributions are as follows:(1)To address the problems of blurred details and hard-detected small objects in underwater images,we propose an underwater fish object detection algorithm based on hard sample mining,which effectively achieves a balance between recall and precision through a two-stage detection strategy from coarse to fine.To reduce the impact of fish deformation on detection,a two-stream Faster RCNN detection network is proposed,which innovatively introduces distance-transformed images and uses contour information as auxiliary features to achieve the detection of hard samples.The effectiveness of our method is verified by conducting experiments on the dataset constructed in this thesis.(2)To address the problems of insufficient illumination,uneven light and low contrast in underwater low-light images,a novel Illumination Parsing model for Underwater low-light Image Enchancement(IP-UIE)is proposed.Inspired by the Lambert-Beer law,we propose an Illumination Parsing Network(IP-Net)to estimate the underwater illumination attenuation.A self-referencing paired image generation method based on color transfer is proposed to generate paired training data,which can better simulate the actual underwater illumination conditions.Extensive experiments and object detection experiments on real images prove that the method can improve the performance of high-level vision tasks while enhancing the visual perception effect.(3)We propose a bilinear network based fine-grained fish recognition algorithm to address the problem of high inter-class similarity and high intra-class variability in fish recognition.By fusing two ways of features through bilinear operation,an effective feature embedding can be obtained,which contributes to capture the relationship between different semantic features simultaneously.In addition,to futher expand the inter-class variability in the feature space and improve the intra-class compactness,we introduce the triplet loss and optimize the network together with the cross-entropy classification loss to improve the recognition accuracy.