| 朱硕,张绪康,宾杰,汪宗洋,江蕊.融合双重观察与注意力机制的灰度图像检测算法[J].电子测量与仪器学报,2025,39(7):192-202 |
| 融合双重观察与注意力机制的灰度图像检测算法 |
| Gray image detection algorithm integrating doubleobservation and attention mechanism |
| |
| DOI： |
| 中文关键词: 目标检测 YOLOv8 灰度图像检测 感受野注意力 空洞卷积 |
| 英文关键词:object detection YOLOv8 gray image detection receptive field attention void convolution |
| 基金项目:国家自然科学基金资助项目（52075520）、江苏双创博士基金（JJSSCBS20210871）项目资助 |
|
|
| Author | Institution |
| Zhu Shuo | Jiangsu Province Engineering Research Center of Photonic Devices and System Integration for Communication Sensing Convergence，
Wuxi University, Wuxi 214105,China |
| Zhang Xukang | School of Electronic and Information Engineering，Nanjing University of Information
Science and Technology, Nanjing 210044,China |
| Bin Jie | Jiangsu Province Engineering Research Center of Photonic Devices and System Integration for Communication Sensing Convergence，
Wuxi University, Wuxi 214105,China |
| Wang Zongyang | Wuxi Xiyuan Technology Co.,Ltd., Wuxi 214000,China |
| Jiang Rui | Jiangsu Province Engineering Research Center of Photonic Devices and System Integration for Communication Sensing Convergence，
Wuxi University, Wuxi 214105,China |
|
| 摘要点击次数: 953 |
| 全文下载次数: 549 |
| 中文摘要: |
| 灰度图像由于其单通道构成的限制，导致图像中目标对比度低、特征信息模糊以及缺少颜色信息，因此检测精度低、且检测难度较大。为提升灰度图像检测的准确率，降低误检和漏检率，提出一种融合双重观察与注意力机制的目标检测算法SAC-YOLO。首先，在主干网络中引入变换空洞卷积，将标准卷积层转换为空洞卷积层，并结合全局上下文模块，提升模型在处理不同尺度和复杂度信息的准确性；其次，特征融合部分采用高效多尺度注意力机制，通过编码全局信息来重新校准各通道权重，跨纬度交互捕捉灰度图像中的像素级关系；最后，添加超分辨率重构检测头，内置感受野注意力模块和卷积模块，关注感受野内空间信息，为大尺寸卷积核提供有效注意力权重，使得模型能够更加精确地适应和表达灰度图像中的小目标信息的特征。在NEU-DET数据集中进行对比实验，改进后的YOLOv8算法对于灰度图像信息的识别精度达到79.3%，相较于YOLOv8原始网络提升了3.1%，由可视化实验可以看出，误检漏检问题得到改善。以上实验结果表明，SAC-YOLO检测效果良好，能够实现在灰度图像场景下的高质量检测。 |
| 英文摘要: |
| Owing to the constraint of the single-channel structure of gray images, the target contrast within the image is low, the feature information is indistinct, and the color information is lacking. Hence, the detection accuracy is low and the detection process is arduous. To enhance the accuracy of gray image detection and reduce the rates of false detection and missed detection, an object detection algorithm, SAC-YOLO, combining dual observation and attention mechanism was proposed. Firstly, transform atrous convolution was integrated into the backbone network to convert the standard convolution layer into an atrous convolution layer, and the global context module was combined to enhance the model’s accuracy in processing information of different scales and complexities. Secondly, the feature fusion part employs an efficient multi-scale attention mechanism to recalibrate the weight of each channel by encoding global information and interactively captures the pixel-level relationship in gray images across latitudes. Finally, a super-resolution reconstruction detection head was added, and a receptive field attention module and a convolution module were constructed to focus on the spatial information within the receptive field and provide effective attention weights for the large-size convolution kernel, enabling the model to adapt and represent the characteristics of small target information in gray images more precisely. The comparison experiment in the NEU-DET dataset reveals that the recognition accuracy of the improved YOLOv8 algorithm for gray image information attains 79.3%, which is 3.1% higher than that of the original YOLOv8 network. It can be observed from the visualization experiment that the issue of false detection and missed detection has been alleviated. The above experimental results indicate that SAC-YOLO has an excellent detection effect and can achieve high-quality detection in grayscale image scenarios. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
