|
| 基于改进MobileNetV3 SSD的河道排污口目标检测研究 |
| Target Detection at Sewage Outlets Based on Improved MobileNetV3-SSD |
| |
| DOI： |
| 中文关键词: MobileNetV3-SSD模型 图像处理 深度学习 河道排污口 |
| 英文关键词: MobileNetv3-SSD Image processing Deep learning Sewage outlet |
| 基金项目:国家自然科学基金资助项目（62176150） |
|
| 摘要点击次数: 225 |
| 全文下载次数: 73 |
| 中文摘要: |
| 为实现对水系入河排污口有效、准确的自动检测，提出一种基于改进MobileNetV3-SSD的深度学习模型。在MobileNetV3-SSD模型的基础上，使用K-means聚类算法和遗传算法，对先验框的宽高比进行调整，使得预测框更好地匹配真实框。引入多尺度特征融合模块，提高模型对小排污口的检测能力。引入改进的CBAM注意力模块，减少模型在排污口检测时计算的参数数量。使用可变形卷积替代普通卷积，自适应地捕获不同排污口的形态与尺度信息，提升模型的特征提取能力。实验结果表明，改进后MobileNetV3-SSD模型的平均精度为89.36%，F1分数为91.88%，较改进前分别提升4.83%和5.03%。 |
| 英文摘要: |
| In order to achieve effective and accurate automatic detection at sewage outlets of water system，a deep learning model based on improved MobileNetV3-SSD was proposed. On the basis of MobileNetV3-SSD model， K-means clustering algorithm and genetic algorithm were used to adjust the aspect ratio of the prior bounding box and make the prediction box match the ground truth box better. The introduction of a multi-scale feature fusion module improved the detection ability of the model at small sewage outlets. The introduction of an improved CBAM attention module reduced the number of parameters calculated by the model during the detection at sewage outlet. Deformable convolution was used instead of ordinary convolution to adaptively capture the shape and size of different sewage outlets， and improve the feature extraction ability of the model. The experimental results showed that the average accuracy of the improved MobileNetV3-SSD model was 89.36%， F1 score was 91.88%， which was 4.83% and 5.03% higher than before improvement， respectively. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
