Research Of The Underwater Stereo Matching Method Based On Binocular Vision

Ocean is the largest treasure house of resources on the earth.The full development of its resources is of great strategic significance to the development of the country,and various underwater detection technologies and exploration equipment are the necessary prerequisites.Since the 21^st century,the demand for underwater detection accuracy has been increasing,and the detection scheme has been gradually transformed from water-acoustic（sonar）to water-light（underwater camera）.The water-light detection technology based on structured light can obtain the 3D information of the target by irradiating the light spot with specific coding and demodulation of the reflected light.However,this scheme has some disadvantages such as high requirement of lighting source,expensive equipment and complex demodulation algorithm.The water-light detection technology based on binocular vision can get the 3D information only through two underwater cameras,which has the advantages of high precision,simple system structure and low cost.Unlike images in air,underwater images often suffer from degradation such as color distortion,low contrast and blurred details,which have great influence on underwater observation and measurement.In view of the above problems,this paper studies underwater image quality enhancement and stereo matching technology respectively based on binocular vision technology.The main content and achievements are as follows:（1）To solve the problem of poor underwater image quality,an underwater image enhancement method based on color adaptive balance and image fusion is proposed.Firstly,the color balance image is obtained by red channel compensation and gray world method,then sharpening andγcorrection are performed respectively.Finally,the two images are fused at multiple scales by image pyramid to obtain the enhanced underwater image.Compared with other classical algorithms,this method has strong adaptability to environment and obvious enhancement effect.（2）To solve the problem of poor accuracy of binocular stereo matching algorithm AD-Census in parallax discontinuity and weak texture region,an improved AD-Census algorithm based on two-stage adaptive optimization and fusion gradient is proposed.The method adaptively adjusts the shape and size of the aggregation window and the weight of the cost fusion according to the pixel position and the length of the cross arm.In addition to absolute difference cost and census cost,pixel gradient information is also added in cost calculation.Compared with the original algorithm,the error of this method is reduced by more than 30%.（3）To solve the problem that traditional stereo matching algorithm can not make full use of global image information and there are gaps in disparity map,an end-to-end stereo matching network base on efficient channel attention module is proposed.Based on the adaptive cost aggregation network AANet,an efficient attentional mechanism module is added to enhance the ability of feature representation to image information,which further improves the accuracy of stereo matching deep learning network.Compared with the original network,the parallax average error of the improved network is reduced by about 9.2%.