基于改进YOLOv7声光融合水下目标检测方法葛慧林，戴跃伟，朱志宇，王彪(江苏科技大学海洋学院，江苏镇江212003)摘要:多变的光照条件及天气状况将会严重影响水下光学图像的成像质量，为提升水下目标检测的稳定性及检测精度，基于深度神经网络模型，对结合光学图像和声呐图形的多模态方法进行研究。首先，针对实时神经网络检测器架构YOLOv7，通过改进该检测器，使其适用于多模态输入。其次，为了有效地结合来自不同模态的影响特征，提出全新的融合模型YOLOv7-Fusion，并通过引入CE-Fusion模块，实现融合效率和准确度的提升。最后，为了解决数据集缺少的问题，利用快速风格和图像处理算法转化的方法，生成人工数据集。所设计的算法及模型目标识别准确率为0.995，具有较高检测精度；Fps为43.4，具有较高处理效率。该模型可支持真实应用，适用于不同类型的水下场景。关键词：改进YOLOv7；水下目标检测；声光融合；光学图像；声呐图像中图分类号：TB566文献标识码：A文章编号：1672–7649(2023)12–0122–06doi：10.3404/j.issn.1672–7619.2023.12.023Researchonacoustic-opticalimagefusionunderwatertargetdetectionmethodbasedonimprovedYOLOv7GEHui-lin,DAIYue-wei,ZHUZhi-yu,WANGBiao(OceanCollege,JiangsuUniversityScienceandTechnology,Zhenjiang212003,China)Abstract:Lightingandweatherconditionsseriouslyaffectthequalityofunderwateropticalimages.Toimprovethestabilityanddetectionaccuracyofunderwatertargetdetection,amulti-modalmethodcombiningopticalimagesandsonargraphicsisstudiedbasedonthedeepneuralnetworkmodel.Firstly,thearchitectureofreal-timeneuralnetworkdetectorYOLOv7isstudied,andthedetectorisimprovedtobesuitableformulti-modeinput.Secondly,inordertoeffectivelycom-binetheinfluencecharacteristicsfromdifferentmodes,YOLOv7-FusionwasproposedandCE-Fusionmodulewasintro-ducedtoimprovefusionefficiencyandaccuracy.Finally,inordertosolvetheproblemofthelackofdataset,faststyleandimageprocessingalgorithmtransformationisusedtogenerateartificialdataset.Thetargetrecognitionaccuracyofthede-signedalgorithmandmodelis0.995andFpsis43.4,withhighdetectionaccuracyandprocessingefficiency.Therefore,themodelcansupportrealapplicationsandissuitablefordifferentunderwaterscenes.Keywords:i...
