水下目标检测在海洋探索、生态保护和水下机器人导航等领域具有重要应用。然而,由于水下环境的复杂性,如光照不均匀、悬浮颗粒干扰和低对比度图像,传统的目标检测方法在水下环境中的表现往往不尽如人意,尤其是面对数据中的噪声问题。为了解决这一问题,本研究提出了一种基于YOLOv7的改进模型用于水下目标检测。我们将YOLOv7作为基线模型,针对其在水下环境中的不足之处,对模型的关键模块进行了优化。具体而言,我们提出了一种漩涡聚合网络模块来破坏噪声数据,并在此过程前引入了空间注意力机制,帮助网络更好地关注重要特征,并抑制不相关的噪声;针对下采样过程中可能存在的信息丢失问题,我们提出了空间到深度池化模块(STD-MP),通过将空间特征转换为深度特征,结合最大池化操作完成下采样过程;最后,我们对损失函数进行了优化。实验结果表明,我们的模型相比于基准模型提升了4.2%的mAP。Underwater object detection has important applications in fields such as ocean exploration, ecological protection, and underwater robotics navigation. However, due to the complexity of the underwater environment, including uneven lighting, interference from suspended particles, and low-contrast images, traditional object detection methods often perform suboptimally in underwater scenarios, particularly when dealing with noisy data. To address this issue, this study proposes an improved model based on YOLOv7 for underwater object detection. We use YOLOv7 as the baseline model and optimize its key modules to overcome its limitations in underwater environments. Specifically, we introduce a vortex aggregation network module to disrupt noisy data, incorporating a spatial attention mechanism before this process to help the network better focus on important features and suppress irrelevant noise. To tackle the issue of potential information loss during downsampling, we propose the Space-To-Depth Pooling (STD-MP) module, which converts spatial features into depth features and combines them with max pooling for downsampling. Finally, we optimize the loss function. Experimental results show that our model achieves a 4.2% improvement in mAP compared to the baseline model.
针对水下环境存在的色偏、细节模糊以及水下目标存在多尺度和遮挡等问题,提出了一种改进YOLOv8s的水下目标检测算法SLG-YOLOv8s。首先,通过Shallow-UWnet网络对水下图像进行增强,提高图像的对比度和清晰度;其次,提出了轻量化多尺度全局注意力(lightweight multi-scale global attention,LMGA)模块,将此模块与YOLOv8s主干部分的C2f融合,通过动态调整权重进行特征重标定,加强特征表达能力,减少计算量;最后,通过聚集和分发(gather-and-distribute,GD)机制采用跨层特征融合的方式增强中间层的信息融合能力,从而提升模型对多尺度目标的检测效率。实验结果表明,SLG-YOLOv8s算法的平均精度均值达到了95.1%,与YOLOv8s算法相比平均精度均值提升了5.1%,精确率和召回率分别提升了4.5%和5.2%,IoU为0.5时海参、海胆、海星、扇贝检测的平均精确率分别提升了4.3%、5.3%、5.6%、5.5%,研究结果可为水下机器人精准捕捞提供重要依据。
水下目标检测由于光的折射与环境杂质等问题,图像质量较差,而检测目标通常较小,且目标间的存在相互遮挡,如鱼、海胆等,使得特征提取困难,检测精度低。而水下终端设备的算力有限,水下作业实时性需求高,限制了深度大模型在水下环境中的应用。提出一种基于Yolov8的水下目标检测模型。为实现模型轻量化同时兼顾保留深层语义特征,首先设计残差分离(DP)模块,引入深度可分离卷积,增加跳跃连接,减少分离计算丢失的特征信息。为了关注重点的特征区域,设计多视角注意力(MA)模块,在通道域复合池化,得到通道注意力权重;在空间域,考虑水下小目标易受噪声干扰特点,使用局部像素注意力代表该区域注意力权重,最后进行特征融合以得到含纹理与背景信息的特征权重。实验证明在水下目标检测竞赛数据集URPC(Underwater Robot Professional Contest)上,改进后的模型精度较原模型提升约7.9%,模型规模减少约20%。
暂无评论