Research On Sunken Ship Identification Method Based On Multi-beam Water Column Image

Based on the national conditions of accelerating the construction of modernization in China,and in view of the development of marine industry and marine surveying and mapping industry,the recognition of underwater targets and the classification of water body images are also gradually increasing in the importance of China’s shipping industry.Compared with other images,sonar images are easily affected by the complex marine environment.The target image has many problems,such as noise snowflakes,large distortion,high similarity with other target objects,and easy to be confused.The multi-beam system can accurately and quickly measure the size,shape and height changes of underwater targets within a certain width along the route,and more reliably depict the three-dimensional characteristics of the seabed terrain.Combined with the navigation positioning and attitude data collected in the field,the multi-beam water column image with high accuracy and resolution is drawn.Therefore,this paper has carried out the research on the sunken ship recognition method based on the multi-beam water column image,so that its accuracy and recognition rate are significantly improved on the original sunken ship image recognition.The main work of this paper is as follows:(1)The Adaboost sunken ship recognition method based on fractal texture features is proposed,and the effect of the calculation of similarity dimension,box dimension and multifractal spectrum used to describe the image texture features is analyzed and compared.Compared with the target recognition results of the traditional multi-beam water column image,it is considered that the accuracy of underwater target recognition in this method is higher,The effect is remarkable in target recognition.(2)A method of sunken ship recognition based on multifractal theory image enhancement and convolution neural network classification is proposed.The convolution neural network is used to identify the multi-beam water column image including the sunken ship.For the training of convolution neural network of sunken and non-sunken target samples,an image enhancement method based on multifractal theory is proposed,which greatly improves the recognition results of convolution neural network.At the same time,the traditional image enhancement methods such as Gaussian filter and median filter are used to carry out comparative experiments,and it is proved that the convolution neural network based on multifractal image data enhancement is superior to the two in accuracy and recognition rate.(3)The full image recognition of the sunken ship based on the multi-beam water column image,respectively,combined with the multi-fractal spectrum texture feature and the convolution neural network,is carried out.Non-maximum suppression method(NMS)is used to avoid the problem of multiple recognition of the same sunken ship target,and the false alarm rate of the sunken ship image is reduced by re-training the false recognition image.In the experiment,the Adaboost sunken ship recognition method based on multifractal spectrum texture features achieved 82.5% of F1 value,while the convolution neural network sunken ship recognition method based on multifractal image enhancement achieved 92.3% of F1 value,which proved that the two sunken ship recognition methods proposed in this paper can effectively identify the sunken ship targets on the multi-beam water column image,and the Adaboost sunken ship recognition method based on multifractal spectrum texture features has high efficiency,The convolution neural network method has a good effect in the identification of shipwrecks,and it has strong operability in practical experiments.Figure [59] table [16] reference [63]...