PDN-YOLO：面向触控屏组件表面缺陷的检测算法

PDN-YOLO: A Detection Algorithm for Surface Defects on Touch Screen Components

摘要: 针对触控屏组件表面缺陷尺度较小、对比度低且易与柔性电路复杂背景相混淆，导致现有检测方法难以兼顾精度与实时性的问题，本文提出了一种基于改进YOLOv10s架构的轻量化高保真检测框架PDN-YOLO。首先，将颈部特征融合网络中的传统卷积模块替换为部分卷积PConv，在保留深层感受野的同时降低计算量，提升了检测效率。其次，采用动态上采样模块DySample替换传统的静态插值算子，该模块利用空间自适应采样机制，改善了微小目标在上采样过程中的边缘模糊问题。此外，引入归一化Wasserstein距离NWD损失函数优化边界框度量机制，缓解了传统损失函数对微小位置偏差过于敏感的问题，提升了定位精度。在包含泡棉、粘膜纸、FPC连接器等8类典型缺陷的1510张触控屏产线数据集上进行的实验结果表明，与基线模型YOLOv10s相比，PDN-YOLO的mAP@[0.5:0.95]指标提升了2.1个百分点，达到87.1%。同时，单图检测速度由6.9 ms缩短至5.9 ms，检测速度提升约14.5%。

Abstract: To address the problem that existing detection methods struggle to balance accuracy and real-time performance due to the small scale, low contrast of surface defects on touch screen components and their easy confusion with the complex background of flexible circuits, this paper proposes a lightweight high-fidelity detection framework PDN-YOLO based on the improved YOLOv10s architecture. Firstly, the traditional convolution modules in the neck feature fusion network are replaced with Partial Convolution (PConv), which reduces the computational load while preserving the deep receptive field and improves detection efficiency. Secondly, the dynamic upsampling module DySample is adopted to replace the traditional static interpolation operator. This module utilizes a spatially adaptive sampling mechanism to alleviate the edge blurring problem of tiny targets during the upsampling process. Furthermore, the Normalized Wasserstein Distance (NWD) loss function is introduced to optimize the bounding box measurement mechanism, which mitigates the excessive sensitivity of traditional loss functions to minor positional deviations and enhances localization accuracy. Experimental results on a dataset of 1510 touch screen production line images containing 8 typical defects such as foam, adhesive film, and FPC connectors show that compared with the baseline model YOLOv10s, the mAP@[0.5:0.95] metric of PDN-YOLO is improved by 2.1 percentage points, reaching 87.1%. Meanwhile, the single-image detection latency is reduced from 6.9 ms to 5.9 ms, representing an approximately 14.5% improvement in detection speed.