| 王新,贾尤杰.结合低光照增强的轻量级瓷砖表面缺陷检测[J].高技术通讯(中文),2026,36(3):318~330 |
| 结合低光照增强的轻量级瓷砖表面缺陷检测 |
| Lightweight tile surface defect detection combined with low-light enhancement |
| |
| DOI：10. 3772 / j. issn. 1002 - 0470. 2026. 03. 010 |
| 中文关键词: 低光照； 降噪； 轻量级； 看一次第8版小模型； 检测头 |
| 英文关键词: low-light, denoising, lightweight, you only look once version 8 small, detection head |
| 基金项目: |
| 作者 | 单位 | | 王新 | (河南理工大学物理与电子信息学院焦作 454000) | | 贾尤杰 | |
|
| 摘要点击次数: 35 |
| 全文下载次数: 43 |
| 中文摘要: |
| 针对瓷砖表面缺陷检测过程中低光照环境导致图像质量退化从而降低检测网络感知能力的问题，本文提出了一种结合低光照图像增强与改进看一次第8版小模型（you only look once version 8 small，YOLOv8s）网络的轻量级瓷砖表面缺陷检测算法。首先，设计了一种无监督的低光照瓷砖图像增强网络，以提高低光照瓷砖图像的质量。该网络通过自校准照明估计提高图像亮度，并通过联合降噪在照明估计之前和之后去除图像噪声。然后，对YOLOv8s网络进行轻量化改进，设计融合重参数化结构与高效多尺度注意力的跨阶段部分残差连接模块(cross stage partial fast module integrated with repvit block and efficient multi，C2f-RVB-EMA)模块与轻量级共享深度分离卷积检测头(lightweight shared depthwise separable convolution detectionhead，LSDSCD)检测头，以增强检测模型的特征判别能力，同时减少计算成本，便于模型部署。实验结果表明，与未进行图像增强的YOLOv8s模型相比，本文所提算法使mAP@50指标提高了4.1%，计算量减少了47.0%，证明了所提方法的有效性。 |
| 英文摘要: |
| To address the issue of image quality degradation under low-light conditions during tile surface defect detection, which reduces the perception capability of detection networks, this paper proposes a lightweight tile surface defect detection algorithm that combines low-light image enhancement with an improved you only look once version 8 small (YOLOv8s) network. First, an unsupervised low-light tile image enhancement network is designed to improve the quality of low-light tile images. This network enhances image brightness through self-calibrated illumination estimation and removes image noise before and after illumination estimation through joint denoising. Then, the YOLOv8s network is lightweighted by designing the cross stage partial fast module integrated with repvit block and efficient multi (C2f-RVB-EMA) module and lightweight shared depthwise separable convolution detection head(LSDSCD) detection head to enhance the feature discrimination ability of the detection model while reducing computational costs, facilitating model deployment. Experimental results show that compared with the original YOLOv8s model without image enhancement, the proposed algorithm improves the mAP@50 metric by 4.1% and reduces the floating-point operations per second (FLOPS) by 47.0%, demonstrating the effectiveness of the proposed method. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
