| 周华平,宋明龙,孙克雷.一种轻量化的水下目标检测算法SG-Det[J].光电子激光,2023,34(2):156~165 |
| 一种轻量化的水下目标检测算法SG-Det |
| SG-Det:A lightweight underwater image target detection method |
| 投稿时间：2022-04-19 修订日期：2022-05-02 |
| DOI： |
| 中文关键词: 神经网络 水下目标检测 轻量化 跨尺度特征增强 |
| 英文关键词:neural networks underwater object detection lightweight networks feature enh ancement by cross-scale |
| 基金项目:安徽省重点研发计划、国际科技合作专项、矿用防爆纯电动运输车辆关键技术的 研究与开发(202004b11010029)资助项目 |
|
| 摘要点击次数: 386 |
| 全文下载次数: 12 |
| 中文摘要: |
| 基于深度学习的目标检测算法在水下进行检测主 要存在两个困难:水下设备的存储和计算能力 有限; 水下图像模糊且小生物聚集。这种局限性要求水下目标检测算法要做到轻量且高 效,因此现有的 目标检测算法不能完全满足水下目标检测的需求。为此本文在SSD(single shot MultiBox d etector) 的网络框架上进行改进,设计了一种轻量化的一阶段检测模型SG-Det。一方面,借鉴Ghos tNet的思想, 对ShuffleNetV2网络进行了重构,提出一种新的轻量化特征提取网络SGnet。此网络进一 步减少了模型参 数量,使模型大小更适合部署在水下设备。另一方面,网络主要是利用6个不同尺度的特征图检 测不同大小 的生物,为此设计了一种引入双分支注意力机制的 跨尺度特征融合模块(cross-scale feature fusion module, AFF)。模块首先引入注意力机制对 输入的特征在 全局通道和局部通道两方面进行加权,突出有用信息,从而减少背景等无关信息的干扰。然 后选取非线 性化程度更高的第4层分别增强前3层的语义信息,以较小的代价使前3层在识别小物体方 面有更好的 表现。模型在中国水下机器人大赛提供的水下数据集UPRC进行试验,平均检测精度(mAP )和速度分 别达到了71.75 FPS,且模型参数量仅有 4.91 M。结果表明,所提出的方法在精度、速度和参数量之间取得了 很好的平衡。 |
| 英文摘要: |
| Deep learn-based object detection algorithm has two main difficulties in underwater target detection,which are limited storage and computing capabilities of underwater equipment and fuzzy underwater images and small organisms gathering.This limitation requires the un derwater target detection algorithm to be lightweight and efficient,so the existing target dete ction algorithm can not fully meet the needs of underwater target detection.In this paper,a lightweigh t one-stage detection model SG-DET is designed to improve the network framework of single shot MultiBox detector (SSD).On the one ha nd,using the idea of GhostNet for reference,ShuffleNetV2 network is reconstructed,and a new lightweight feature extraction network SGnet is proposed,which further reduce the number of model parameters and made the model size more suitable for deployment in underwater eq uipment.On the other hand,the network mainly uses six feature maps of different scales to dete ct organisms of different sizes.For this purpose,a cross-scale feature fusion module (AFF) using two-branch attention mechanism is designed. The AFF module firstly introduces the attention mechanism to weight the input fe atures in both global channel and local channel,so as to highlight useful information.To impr ove the problem of low detection accuracy caused by small and fuzzy objects,the fourth layer with a higher degree of nonlinearity is selected to enhance the semantic information of the first three layers,so that the first three layers have a better performance in identifying small objects with a small cost.The model is tested on the underwater data set UPRC provided by China Underwater Robot Compet ition,and the mAP and speed reach 71.75% and 69 FPS respectively,while the number of model p arameters is only 4.91 M.The results show that the proposed method achieves a good balance be tween precision,speed and number of parameters. |
| 查看全文 下载PDF阅读器 |
| 关闭 |
|
|
|
|
|