Research On Underwater Object Perception Method For Side-Scan Sonar Image

Underwater object perception has become one of the most prevalent research topics in marine exploration,with an urgent demand for prospecting and development of unknown ocean.Side-scan sonar has been widely operated in underwater experiments which requires larger-scale and long-range perceptive ability due to its wide sensing range and high detection accuracy.However,existing underwater object perception methods for side-scan sonar image encounter multiple challenges,such as complex and changing marine environment,a large amount of redundant data and a dilemma that a significant amount of data cannot be shared publicly to support scientific research for its sensitivity,which results in a compromising situation between perception efficiency and model plasticity.This paper proposed a method for side-scan sonar image to comprehensively improve efficiency and plasticity of object perception.The object perception method consists of two parts,including object segmentation and target recognition,which are as follows.In the first stage,object segmentation algorithm is required to balance a great tradeoff between high detection accuracy and low false alarm rate for ensuring that marine exploration can be operated robustly in unknown ocean without any target characteristics or datasets.This paper proposed an object segmentation method based on constant false alarm rate,which can achieve the following two functions.(1)Robust locating.To get rid of dependence of datasets,the method is designed mathematically to locate object regions in a large amount of data.(2)Accurate detection.The method incorporates shadow information of objects to improve detection accuracy when facing both large and small objects.Moreover,to be suitable for scenarios where the number of objects in an image is unknown,pixel homogeneity is introduced to this algorithm for getting rid of dependence of detection results on parameter setting.Comparative experiments have illustrated that the segmentation algorithm outperforms well.In the second stage,target recognition performance is mostly influenced by sidescan sonar image datasets,however,the datasets cannot effectively support numerous and unknown marine activities because of their poor sharing,small sample size and imbalanced categories.This paper proposed a class-incremental learning method for side-scan sonar images,which utilized distillation techniques for reviewing old class knowledge and introduces focal loss to alleviate the problem about sample imbalance,enabling continuous learning is processed with only a small amount of old data.Comparative experiments have shown that this class-incremental method effectively cracks the hard nuts concerning catastrophic forgetting and data imbalance.The experimental results demonstrate that this object perception method can achieve a great tradeoff between high detection accuracy and low false alarm rate in object segmentation,as well as support continuous learning for target recognition with imbalanced samples.