Research On Object Detection Of Sonar Image Based On Foreground Enhancement

Due to the complexity of the underwater environment and the special imaging mechanism of the sonar system,sonar images are quite different from natural images.First,the noise in the sonar image is very large and complex,and secondly,the image contains less high-frequency information,mainly including edges and details,so the quality of the sonar image is not clear enough.Existing target detection methods are mainly aimed at optical images and do not consider these two characteristics of sonar images,resulting in low detection performance.For this reason,based on the two characteristics of noise and high-frequency information loss that widely exist in sonar images,this paper has carried out the following three research work:(1)Aiming at removing the multiplicative speckle noise with deep influence in sonar images,a sonar image noise reduction network based on self-supervised learning is designed.Specifically,starting from the imaging mechanism of the sonar system,the noise existing in the image is analyzed in detail,and the corresponding noise model is constructed.Next,since the noise reduction network based on self-supervised learning can only deal with additive noise in theory,it is necessary to perform logarithmic transformation on the noisy image to convert the multiplicative speckle noise in the image into an additive noise model.Then,the transformed image is input into the noise reduction network for processing,and the image is restored by exponential transformation after noise reduction.In the denoising network,in order to make the output image contain detailed information as much as possible,a detail measurement loss item is added to the corresponding loss function.The experimental results show that the introduction of the loss term effectively improves the performance of the denoising network,and has a good speckle noise suppression effect.(2)Aiming at removing the widespread noise in sonar images,a semantic enhancement module for the foreground of sonar images is designed.Specifically,a feature fusion method is used to generate a strong semantic feature map,which has strong location and semantic information at the same time.Then,the relationship between the modeled strong semantic feature map and each input feature map is displayed,and all input feature maps are enhanced with the corresponding relationship matrix.In the foreground semantic enhancement module,techniques such as feature fusion and receptive field module are used,and the resulting feature maps are more discriminative.In sonar images,the background is more complicated due to the existence of noise,and the background point is regarded as the foreground point and the prediction frame is generated during the target detection process.After foreground semantic enhancement,the gap between foreground points and background points in the feature map will be enlarged,and the invalid prediction frame generated by background points will be suppressed.The experimental results show that the foreground semantic enhancement module can suppress the noise point features,and the enhanced features can improve the subsequent target detection ability.(3)Aiming at the lack of high-frequency information in sonar images,a foreground edge enhancement module for sonar images is designed.And using the method of adaptive feature fusion,the features from the foreground semantic enhancement module and the foreground edge enhancement module of the sonar image are fused,and the sonar image foreground enhancement network is designed.Specifically,starting from the edge characteristics of the image,it is analyzed that the spatial context information in different directions has a high effect on edge recognition.Then,a foreground edge enhancement module is constructed according to these characteristics of the edge: using recurrent neural network to extract the spatial context information of the four cardinal directions of each pixel,and assigning different weights to this information through the spatial attention mechanism.After edge enhancement,the features of edge points in the feature map are enhanced due to the obtained spatial context information.At the same time,due to the working principle of the recurrent neural network,the spatial context information continues to propagate to the inside of the edge,so the foreground points inside the edge are also enhanced.Finally,through experiments,the effectiveness of the foreground edge enhancement module and the sonar image foreground enhancement network is proved.