基于改进Deformable-DETR的水下图像目标检测方法  

An object detection method of underwater image based on improved Deformable-DETR

作　　者：崔颖[1] 韩佳成 高山[1] 陈立伟[1] CUI Ying;HAN Jiacheng;GAO Shan;CHEN Liwei(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001

出　　处：《应用科技》2024年第1期30-36,91,共8页Applied Science and Technology

基　　金：黑龙江省自然科学基金项目(LH2020F021)。

摘　　要：针对由于水下复杂环境造成的目标检测效果较差、检测精度较低的问题,基于Deformable-DETR算法提出一种改进的水下目标检测算法Deformable-DETR-DA。使用空间注意力模块结合标准Transformer块设计了一个用于增加模型深度的深度特征金字塔(deep feature pyramid networks,DFPN)模块,将其嵌入到模型中提高模型对深层纹理信息的提取能力。使用注意力引导的方式对原模型中编码器部分进行改进,加强了对特征信息的聚合能力,提高了模型在复杂环境下的检测能力。针对URPC数据集,模型各交并比尺度的平均准确度(average precision,AP)为39.5%,相比原模型提升1%,与一些DETR(detection transformer)类的模型相比,不同目标尺度的平均准确度均有1%~4%左右的提高,表明改进的模型能够很好解决复杂环境的水下目标检测的问题。本文提出的模型可作为其他水下目标检测模型设计的参考。Aiming at the problem of poor object detection effect and low detection accuracy caused by complex underwater environments,an improved underwater target detection algorithm Deformable-DETR-DA is proposed based on the Deformable-DETR algorithm.Using the spatial attention module and the standard Transformer block,a DFPN block is designed to increase the depth of model,and the DFPN block is embedded into the model to improve the ability of the model to extract the deep texture information.The encoder part of the original model is improved by using attention guidance,which strengthens the aggregation ability of feature information and improves the detection ability of the model in a complex environment.For the URPC dataset,the average precision(AP)of each intersection over union scale of the model is 39.5%,which is 1%higher than the original model.Compared with some DETR-like models,the average precision of different object scales is improved by 1%~4%,which shows that the improved model can well solve the problem of underwater object detection in complex environments.The model proposed in this paper can serve as a reference for the design of other underwater object detection models.

关 键 词：水下光学图像 Deformable-DETR 目标检测 TRANSFORMER 注意力机制 深度学习 图像处理 残差网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]