感知增强混合网络的水下目标检测

打开文本图片集

中图分类号：TP391 文献标识码：Adoi：10.37188/0PE.20253308.1303

Abstract：Underwater object detection technology plays an important role in areas of marine resource exploration and environmental protection. However，the problems of blurred imaging and variable object scales in underwater environments pose dificulties fordetection tasks. As a result，it is challenging for accurate underwater object feature extraction，which influences the detection performance of existing methods.To solve the above-mentioned problem，a feature enhanced hybrid network was proposed to improve the detection accuracy of underwater objects.Firstly，a global-local hybrid feature enhancement network was constructed.The long-range global information in the image was extracted via self-attention mechanisms，and the richer localdetailed information was further calculated through the devised convolutional attention enhancement module.The global and local relationships in the images could be beter established， hence the multiscale feature representation powers of the network were enhanced. Subsequently，in order to suppress the interference of imaging blurriness and low contrast in underwater environments，a twostage object perception enhanced detection head was constructed. The depth of the first-stage region proposal network was increased，thus more semantic information of underwater objects could be extracted. Besides，the self-atention mechanism was introduced in the second stage to suppress the interference from background noise.Moreover，an intersection-over-union branch was added to further integratethe prior information of objects obtained from the first stage into the second stage.The proposed method achieves （204号 37.8% ， 61.8% ，and 82.0% ， 98.9% of mAP0.5：0.95 and AP50 on the TrashCan and WPBB datasets respectively. The qualitative and quantitative comparison experimental results demonstrate that this method could obtain superior detection results for various underwater objects.

Key Words：underwater object detection；feature enhancement；self-attention；hybrid network

1引言

利用水下目标检测技术可以自动定位出水中感兴趣目标在图像中的位置，并给出目标的类别信息。（剩余14372字）