Deep Learning Recognition Algorithm For Marine Benthos

Marine benthos such as sea cucumbers,sea urchins and scallops are tasty and expensive and are the main objects of fisheries.Traditional fishing for marine benthos relies on divers entering the water,and as a result,divers are exposed to very high risks,and even if they complete their work safely,they are prone to diving sickness over long periods of time.With advances in technology,underwater robots have greatly improved the efficiency of seafood harvesting,but for many reasons,including complex underwater environments,underwater robots still have difficulties in performing underwater object recognition and harvesting tasks.In this thesis,four common marine benthos are studied and corresponding solutions are proposed for the multifaceted problems of underwater imagery.The main research elements of this thesis are as follows:Firstly,this thesis reproduces a variety of underwater image enhancement algorithms based on the principle of underwater image imaging,on the basis of which a variety of image enhancement algorithms are combined to solve the problems of colour bias,blurring and low contrast of underwater images respectively,making a lot of work for subsequent algorithm comparison.Meanwhile,this thesis proposes a depth estimation network-based colour reconstruction algorithm for underwater images based on a novel underwater imaging model.The algorithm first trains the depth estimation network through transfer learning,and then uses the novel underwater imaging model to complete the colour reconstruction of underwater images,which well solves the problem of missing key parameters in the dataset and not only corrects the colours of the images,but also restores the details of the images.Secondly,by analysing the difficult points of the underwater recognition task,this thesis proposes an underwater object recognition algorithm based on an attention mechanism.The algorithm uses a combination of feature fusion module and feature enhancement module to adaptively extract multi-scale feature maps using convolutional neural networks,while a cascaded attention mechanism scheme consisting of an anchor refinement module,a spatial attention module and a object recognition module is used to improve the classification performance and regression performance of underwater objects,and its recognition results provide subsequent control of the robot.The multi-modular structure obtains multi-scale contextual features and the cascaded attention mechanism scheme highlights objects in spatial regions on a given feature map,making the overall network more conducive to identifying small underwater objects.The algorithm improves the recognition accuracy while taking into account the recognition speed.Finally,this thesis designs and implements a ROS functional package for marine benthos creature recognition.On an underwater robot simulation platform,this thesis encapsulates the colour reconstruction algorithm and object recognition algorithm into a ROS function package for real-time communication.Experiments prove that the function package can handle both algorithms in parallel and has the characteristics of accuracy,real-time and high compatibility.