Research On Segmentation Algorithm Of Sea Target In Complex Weather Based On Deep Learning

As a major marine country,China has a great potential for marine economic development,but there are many disputes over maritime rights and interests.The disputes on the sea surface from time to time pose a threat to our territorial sovereignty.Chinese maritime security is facing a more severe test,so it is of great significance to strengthen Chinese sea surveillance and sea defense.The sea target segmentation technology plays an important role in the field of sea surface monitoring.In addition,in the scene of unmanned war on the sea,the response speed and hit rate of the ship-borne weapon can be greatly improved by improving the sea target segmentation ability of the ship-borne visual investigation system,which can further enhance our country’s sea surface defense capability.Sea fog weather is a common disturbance in the navigation of ships,which will adversely affect the target segmentation task.Therefore,combining the advantages of deep learning in the learning of large data features,this paper presents a sea target segmentation algorithm in complex weather based on deep learning.First of all,an improved dark channel prior algorithm is proposed to preprocess the sea target image to solve the problem that the sea fog weather will affect the segmentation results.Based on the description of the atmospheric scattering model,the defogging principle and derivation process of the dark channel prior algorithm are explored in detail.Aiming at the shortcomings of too much time consuming of the original dark channel algorithm and inaccurate estimation of atmospheric light,an acceleration strategy based on downsampling and an improved algorithm for atmospheric light estimation optimization are proposed.Through the comparative experiment of the defogging algorithm and the evaluation of the defogging effect,it is proved that the improved dark channel defogging algorithm has a better defogging effect.Secondly,the Deep Labv3+ semantic segmentation algorithm is selected to segment the sea target.In order to solve the problem that the target segmentation accuracy of the original Deep Labv3+ network is not high enough,an improved Deep Labv3+ network model is proposed.The main improvement is that the optimal DPC architecture is used to replace the ASPP module in the original Deep Labv3+ network,and the depthwise separable convolution is used in the decoder.The experimental results show that the accuracy of the improved Deep Labv3+ network in the task of sea target segmentation is higher than that of the unimproved Deep Labv3+ network,and its overall MIo U index is improved by 4.1%.Finally,the Mask R-CNN network is selected to segment the sea target,which solves the problem of target conglutination in segmentation results under occlusion condition existing in the improved Deep Labv3+ algorithm.Aiming at the disadvantage of slow segmentation speed in the network,a network acceleration improvement based on the fusion of BN layers to convolution layers is proposed,and the loss function is improved to optimize the segmentation effect of the target boundary.The experimental results show that the speed and recognition rate of the improved Mask R-CNN network are improved.After that,a comparative experiment is set up,the environment includes single target,multi-target,occluded target and small target under different background,and the influence of sea fog weather on the segmentation result is considered in the experiment.The experimental results show that the improved Deep Labv3+ network and the improved Mask R-CNN network have their own advantages in segmentation speed and accuracy respectively.