Research On Sonar Image Target Recognition Method Based On Spatial Position Information And Machine Learning

Scientific development of marine resources is of great significance to economic development and environmental protection,and underwater sonar imagery has received extensive attention as an important means of marine research.With the introduction of deep learning modules,underwater sonar image processing technology is becoming more and more mature,but the sonar imaging process is susceptible to problems such as low sonar image quality and weak contrast caused by environmental interference,which greatly restricts the accuracy of sonar image target recognition algorithm.In addition,because the traditional methods of image preprocessing and height information restoration do not fully consider the characteristics of sonar images,there are great limitations in improving the accuracy of sonar image target recognition.Therefore,this thesis conducts detailed research on the preprocessing,height information restoration and target recognition process of sonar image recognition,and the main contents are as follows:(1)Preprocessing process for sonar image recognition.Since sonar images have the characteristics of missing edge information and blurred texture,a sonar semantic enhancement method based on improved Lee filtering is proposed to preprocess sonar sub-graphs,including: 1)the traditional image processing process uses the histogram linear transformation method to enhance the image of the sonar subgraph from the pixel value level on the basis of grayscale and missing information interpolation processing;2)By improving the Lee filtering method,the sonar semantics in the sonar subgraph are weakly recognized,and the local sonar semantics are enhanced according to the recognition results;3)The proposed sonar semantic enhancement method based on improved Lee filtering is compared with the traditional image preprocessing results,and the good actual performance of the method is verified.(2)Contour information extraction process for sonar image recognition.In order to solve the problem of target contour blurring in sonar image recognition,an underwater sonar image matching method based on improved K-Means clustering is proposed to restore the edge information and further sharpen the target contour,including: 1)extracting the target feature points based on the gradient threshold,establishing the scene segmentation boundary of the target,shadow and background,so as to obtain three grayscale image regions with different brightness with sudden edge pixel values;2)A method based on improved K-Means clustering is proposed to calculate and represent the target area and Gaussian distribution,and eliminate the misidentified small targets to obtain the position of the target trailing edge point;3)According to the scene analysis conclusion,establish the position relationship between the target and the shadow,and use the method of obtaining the similarity of the target trailing edge point to mark the contour information of the target front-end point and the shadow trailing edge point.(3)Three-dimensional reconstruction process for sonar image recognition.In order to further restore the height information of the three-dimensional target in the two-dimensional image to complete the three-dimensional information acquisition,a three-dimensional reconstruction method based on the pitch angle calculation of sonar image is proposed,which includes: 1)according to the construction results of the physical imaging model of sonar image,the calculation of seabed plane normal vector is used to obtain the accurate pitch angle of sonar image;2)According to the normal vector and the pitch angle calculation results of the sonar image,the height of the objects at different heights in the sonar image is identified to strengthen the height information in the image;3)Aiming at the pose noise problem and parameter error problem in the preliminary height information,a three-dimensional map coordinate generation method based on pose estimation and parameter optimization is proposed,and the three-dimensional information results based on the emitter of sonar image recognition are obtained.(4)Target recognition process for sonar image recognition.A transfer learning convolutional neural network based on multi-modal and multi-framework is proposed,which includes: 1)In order to further solve the training overfitting problem that is easy to occur in typical sonar image recognition methods,the idea of transfer learning is introduced into classical networks,which reduces the weight calculation burden of the network coding part,and improves the anti-fitting robustness and recognition accuracy of the network.2)Aiming at the recognition accuracy limitation caused by height information that is not considered in the traditional deep learning sonar image target recognition method,a multimodal neural network is used to integrate the height information into the neural network as an independent modal to guide the recognition of sonar images;3)Aiming at the portability problem of deep learning network construction under different frameworks,multi-base frameworks are used to verify and build separately.Experimental analysis shows that the proposed convolutional neural network based on multimodal transfer learning has high feasibility,and obtains a recognition accuracy much higher than that of traditional recognition methods such as SVM classifier and Alex Net network in actual recognition,which verifies that the proposed model has better recognition ability and generalization ability.In addition,the construction of noise models,three-dimensional information reconstruction methods and object recognition models still need to be studied in more depth to further improve the accuracy of target classification of sonar images.