[1]刘梦菲,毛建华,陆小锋.基于注意力和上下文感知的海面渔桩检测[J].计算机技术与发展,2023,33(08):144-150.[doi:10. 3969 / j. issn. 1673-629X. 2023. 08. 021]
　LIU Meng-fei,MAO Jian-hua,LU Xiao-feng.Detection of Fishing Piles on Sea Surface Based on Attention and Context Awareness[J].,2023,33(08):144-150.[doi:10. 3969 / j. issn. 1673-629X. 2023. 08. 021]

点击复制

基于注意力和上下文感知的海面渔桩检测()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

33

期数:

2023年08期

页码:

144-150

栏目:

人工智能

出版日期:

2023-08-10

文章信息/Info

Title:

Detection of Fishing Piles on Sea Surface Based on Attention and Context Awareness

文章编号:

1673-629X(2023)08-0144-07

作者:

刘梦菲¹ ; 毛建华¹ ; 陆小锋¹; 2

1. 上海大学 通信与信息工程学院,上海 200444;

2. 上海大学 温州研究院,浙江 温州 325000

Author(s):

LIU Meng-fei1 ; MAO Jian-hua1 ; LU Xiao-feng1; 2

1. School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China;

2. Wenzhou Institute of Shanghai University,Wenzhou 325000,China

关键词:

小目标检测; YOLOv5; 上下文信息; 坐标注意力机制; 渔桩

Keywords:

small target detection; YOLOv5; contextual information; coordinate attention mechanism; fishing pile

分类号:

TP391

DOI:

10. 3969 / j. issn. 1673-629X. 2023. 08. 021

摘要:

为了精确打击拦网捕鱼行为,提高渔政执法的效率,将目标检测技术应用于无人机采集的违法渔桩影像。 针对海面渔桩小目标检测精度低以及误检、漏检率高等问题,提出一种 AECA-YOLO( Attention Enhanced Contextual Aware) 模型对渔桩目标进行更精确的检测与定位。 该算法通过连接小目标和上下文信息进行数据增强,在原始 YOLOv5 算法的骨干网络中添加一个坐标注意力机制,加强特征通道间的关联;其次,提出一种嵌入注意力机制的亚像素上采样结构替代最邻近上采样,丰富目标区域的细节信息;同时,采用解耦检测头分离分类与定位过程,提高训练的速度与稳定性;最后,调整网络的定位损失函数,改善位移对小目标的剧烈干扰。 实验结果表明,将改进后的算法应用在海面背景下的渔桩小目标检测中,相比原始 YOLOv5 算法,在检测速度相当的情况下平均检测精度提高了 29. 7% ,召回率提升了 18. 9% ,检测速度为 52. 37 FPS,能够满足实时性的需求。 渔桩小目标的实时高精度检测为智能渔政执法提供了有力的解决方案。

Abstract:

In order to accurately combat interception fishing behavior and improve the efficiency of fisheries enforcement,object detectiontechnology is applied to the images of illegal fishing piles collected by drones. To address the problems of low detection accuracy andhigh false positive and miss rate of fishing piles small targets on the sea surface, an AECA - YOLO ( Attention Enhanced ContextualAware) model is proposed for more accurate detection and localization?

of fishing pile targets. The algorithm performed data enhancementby connecting small targets with contextual information, and added a coordinate attention mechanism to the backbone of the originalYOLOv5 algorithm to strengthen the association between feature channels. Secondly,a sub-pixel upsampling structure embedded in theattention mechanism was proposed to replace the nearest neighbor upsampling to enrich the detail information of target regions.　Meanwhile,the decoupled detection head was used to separate the classification and localization processes to improve the speed andstability of training. Finally, the localization loss function of the network was adjusted to improve the dramatic interference ofdisplacement on small targets. The experimental results show that the improved algorithm applied to the detection of fishing piles in thesea surface background,compared with the original YOLOv5 algorithm, improves the average detection accuracy by 29. 7 % and therecall rate by 18. 9 % with comparable detection speed, the detection speed is 52. 37 FPS, which meets the demand of real - timeperformance. The real-time and high - precision detection of small targets of fishing piles provides a powerful solution for intelligentfishery law enforcement.

常用功能