| 引用本文: | 陈 洋1 ,张晓光1,3,4 ,陆凡凡2 ,束正华2 ,徐文强5 ,王 涵2 ,徐新志2.一种结合上下文感知模块的高压微雾灰尘检测方法(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2025，42（4）：116-121 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 本文已被：浏览 128次 下载 82次 |
码上扫一扫！ |
|
|
| 一种结合上下文感知模块的高压微雾灰尘检测方法 |
|
陈 洋1 ,张晓光1,3,4 ,陆凡凡2 ,束正华2 ,徐文强5 ,王 涵2 ,徐新志2
|
|
1. 上海智质科技有限公司,上海 201801
2. 安徽智质工程技术有限公司,安徽 芜湖 241000
3. 中国科学技术大学 计算机科学与技术学院,合肥 230026
4. 长三角信息智能创新研究院,安徽 芜湖 241000
5. 巢湖海螺水泥有限公司,安徽 巢湖 238000
|
|
| 摘要: |
| 目的 针对堆场下料口灰尘大,喷雾系统无法快速精准定位除尘等问题,提出一种结合上下文感知模块的检
测方法以实现对现场粉尘的有效检测,辅助高压喷雾系统快速除尘。 方法 首先模型的主干网络为轻量级网络
EfficientNetB0,在实现高效特征提取的同时可以大大减少网络的模型参数量,提升部署阶段应用效率;其次利用
CoT( Contextual Transformer) 模块充分探索相邻层级之间的上下文信息,以一种结合静态与动态信息的方式提升自
注意力学习,增强网络特征提取能力,进而提升输出特征的表达能力;最后在 3 个输出层之间进行通道调整与融合
之后输入自适应空间特征融合( Adaptively Spatial Feature Fusion,ASFF) 网络,进一步融合各通道之间的信息特征,
有助于特征细节信息的学习。 结果 整个网络的模型大小为 20. 42 MB,有利于模型的嵌入使用,均值平均精度
( mean Average Precision,mAP) 为 95. 98%。 结论 提出的结合上下文感知模块的检测方法应用于堆场下料口灰尘
检测不仅降低了计算量且在精确度方面有一定优势,满足检测要求。 |
| 关键词: 缺陷检测 CoT 非对称卷积 自适应特征融合 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| A High-pressure Micro-mist Dust Detection Method Incorporating a Context-aware Module |
|
CHEN Yang1 ZHANG Xiaoguang1 3 4 LU Fanfan2 SHU Zhenghua2 XU Wenqiang5 WANG Han2 XU Xinzhi2
|
|
1. Shanghai Zhizhi Technology Co. Ltd. Shanghai 201801 China
2. Anhui Zhizhi Engineering Technology Co. Ltd. Anhui Wuhu 241000 China
3. School of Computer Science and Technology University of Science and Technology of China Hefei 230026 China
4. Yangtze River Delta Information Intelligence Innovation Research Institute Anhui Wuhu 241000 China
5. Chaohu Conch Cement Co. Ltd. Anhui Chaohu 238000 China
|
| Abstract: |
| Objective To address issues such as excessive dust at the material unloading point in the yard and the inability
of spray systems to quickly and accurately locate and remove the dust this paper proposed a detection method integrating
a context-aware module to effectively detect on-site dust and assist the high-pressure spray system in rapid and precise dust
removal. Methods The backbone network of the model was a lightweight network EfficientNetB0 which significantly
reduced model parameters while achieving efficient feature extraction and improving deployment efficiency. Additionally, the CoT Contextual Transformer module was used to explore contextual information between adjacent layers enhancing
self-attention learning with a combination of static and dynamic information to improve feature extraction and expression.
Finally after channel adjustment and fusion between three output layers the input was passed to the Adaptive Spatial
Feature Fusion ASFF network for further integration of information features across channels so as to facilitate the
learning of feature details. Results The total model size of this method is 20. 42MB facilitating model embedding usage
and the mean Average Precision mAP is 95. 98%. Conclusion The proposed context-aware module integrated
detection method reduces computational load and has a certain advantage in accuracy for detecting dust at the material
unloading point in the yard meeting detection requirements effectively. |
| Key words: defect detection CoT asymmetric convolution adaptive feature fusion |
