基于改进CenterNet的水下目标检测算法  

Underwater Object Detection Algorithm Based on Improved CenterNet

作　　者：王蓉蓉 蒋中云[2] Wang Rongrong;Jiang Zhongyun(College of Information,Shanghai Ocean University,Shanghai 201306,China;College of Information Technology,Shanghai Jian Qiao University,Shanghai 201306,China)

机构地区：[1]上海海洋大学信息学院,上海201306 [2]上海建桥学院信息技术学院,上海201306

出　　处：《激光与光电子学进展》2023年第2期229-238,共10页Laser & Optoelectronics Progress

基　　金：上海市属高校应用型本科试点专业建设项目(Z32004-17084);上海市教育委员会一流本科专业建设专项项目(JYLB202002)。

摘　　要：针对常规目标检测器检测水下目标时存在特征提取困难、目标漏检等问题,提出一种改进CenterNet的水下目标检测算法。首先,使用高分辨率人体姿态估计网络HRNet代替CenterNet模型中的Hourglass-104骨干网络,降低模型参数量,提升网络推理速度;其次,引入瓶颈注意力模块,在空间维度及通道维度进行特征增强,使网络关注重要目标特征信息,提高检测精度;最后,构建特征融合模块,融合网络内部丰富的语义信息和空间位置信息,并利用感受野模块增强融合后的特征,提高网络多尺度目标检测能力。在URPU水下目标检测数据集上进行实验,与CenterNet相比,所提算法的检测精度可达77.4%,提升1.5个百分点,检测速度为7 frame/s,提升35.6%,参数量为30.4 MB,压缩84.1%,同时与其他主流目标检测算法相比具有更高的检测精度,在水下目标检测任务上更具优势。Aiming at the problems of conventional detectors in detecting underwater objects,such as difficulty in feature extraction and missing detection of objects,an improved CenterNet underwater object detection method is proposed.First,a high resolution human posture estimation network HRNet is used to replace the Hourglass-104 backbone network in CenterNet model to reduce the amount of parameters and improve the speed of network reasoning;then,the bottleneck attention module is introduced to enhance the features in the spatial and channel dimensions,and improve the detection accuracy;finally,a feature fusion module is constructed to integrate the rich semantic information and spatial location information in the network,the fused features are processed by receptive field block to further improve the multi-scale object detection ability of the network.A comparison experiment is carried out on the URPU underwater object detection dataset.Compared with CenterNet network,the detection accuracy of the proposed algorithm can reach 77.4%,increased by 1.5 percentage points,the detection speed is 7 frame/s,increased by 35.6%,the amount of parameters is 30.4 MB,compressed by 84.1%.Compared with the mainstream object detection algorithm,this algorithm also has higher detection accuracy,which has higher advantages in underwater object detection.

关 键 词：机器视觉 水下目标检测 CenterNet 高分辨率网络 注意力机制 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]