| 陈志旺* **,雷春明*,吕昌昊***,王婷*,彭勇****.基于注意力与密集重参数化的目标检测算法[J].高技术通讯(中文),2024,34(3):233~247 |
| 基于注意力与密集重参数化的目标检测算法 |
| Object detection algorithm based on attention and dense reparameterization |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2024. 03. 002 |
| 中文关键词: 目标检测； 重参数化； 注意力机制； 特征融合； 上采样； 正负样本匹配 |
| 英文关键词: object detection, reparameterization, attention mechanism, feature fusion, upsampling, positive and negative sample matching |
| 基金项目: |
| 作者 | 单位 | | 陈志旺* ** | (*燕山大学智能控制系统与智能装备教育部工程研究中心秦皇岛 066004)
(**燕山大学工业计算机控制工程河北省重点实验室秦皇岛 066004)
(***燕山大学河北省电力电子节能与传动控制重点实验室秦皇岛 066004)
(****燕山大学电气工程学院秦皇岛 066004) | | 雷春明* | | | 吕昌昊*** | | | 王婷* | | | 彭勇**** | |
|
| 摘要点击次数: 674 |
| 全文下载次数: 449 |
| 中文摘要: |
| 针对目标检测任务中背景复杂、目标尺寸差异大等因素导致目标检测结果较差的问题，本文提出基于注意力和密集重参数化的目标检测算法。首先，基于CSP DarkNet提出高效的特征提取网络，主要包括密集重参数化模块和CASA模块2个设计。前者利用密集连接保留浅层特征，又通过重参数化结构降低网络复杂度；后者CASA模块用于获取需要的目标信息。其次，特征融合在特征金字塔(FPN)和路径聚合网络(PAN)的基础上，引入内容感知特征重组（CARAFE）进行上采样，有效解决了邻近插值法等未能捕捉丰富语义信息的问题；提出更高效的C3-G模块，获取丰富的梯度信息，增强模型表达能力和感知能力；同时，引入深度可分离卷积提升运算效率。最后，检测输出采用在更大范围上跨领域正负样本匹配策略扩充正样本数量，提升检测效果。该算法在MS COCO和PASCAL VOC数据集上的mAP@0.5分别达到了57.5%和83.0%，充分说明了本文算法的先进性。 |
| 英文摘要: |
| This paper presents an object detection algorithm using attention and dense reparameterization to tackle challenges posed by complex backgrounds and variations in object sizes, which can adversely affect detection results. The proposed algorithm consists of two key components within an efficient feature extraction network based on CSP-DarkNet: the dense reparameterization module and the coordinate and spatial attention (CASA) module. The former leverages dense connections to retain shallow features while reducing network complexity through reparameterization structures, while the CASA module captures necessary target information. Feature fusion is performed using feature pyramid network （FPN） and path aggregation network（PAN）, and upsampling is achieved through content-aware reassembly of features (CARAFE), addressing the issue of insufficient capture of rich semantic information. To enhance model capabilities, a more efficient C3-G module is introduced to obtain gradient information, and depthwise separable convolution is employed to improve computational efficiency. Lastly, the detection output is enhanced by employing a cross-domain positive-negative sample matching strategy on a larger scale, augmenting positive samples and improving detection performance. Experimental results showcase the algorithm’s advancements, achieving mAP@0.50 scores of 57.5% and 83.0% on the MS COCO and PASCAL VOC datasets, respectively. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
