改进YOLOv8的晶体熔接阶段图像视觉识别检测方法

Improving the Visual Recognition and Detection Method for Crystal Fusion Stage Images in YOLOv8

摘要: 直拉硅晶体生长熔接流程中生长界面温度检测是保障后续引晶成功的重要任务。现有的目标检测模型对熔接凸点目标检测存在定位不准确，误检率高等问题，本文提出了一种改进了基于特征增强的YOLOv8算法。首先，针对熔接光圈图像中小目标容易出现误检和漏检的常见问题，引入BiFPN的思想对YOLOv8m中的颈部部分进行改进。为了进一步提升检测精度，在特征融合网络中采用了更轻量的动态上采样算子DySample，以提高融合特征的质量和丰富度。在工业提供的数据集上评估了YOLOv8-A模型，实验结果表明，与原来算法相比，YOLOv8-A的参数量和计算量分别减少至2.19 × 10^7，同时实现了98.2%的mAP，对小目标的检测提升了5.8个百分点。通过与其它主流目标检测算法比较，验证了该方法的有效性和优越性。

Abstract: The detection of the growth interface temperature in the Czochralski process for silicon crystal growth welding process is considered an important task for ensuring the successful subsequent crystal pulling. Existing target detection models exhibit issues such as inaccurate localization and high false detection rates in the detection of welding bump targets. An improved YOLOv8 algorithm based on feature enhancement is proposed in this paper. Firstly, to address the common problems of false and missed detections of small targets in welding aperture images, the concept of BiFPN is introduced to enhance the neck part of YOLOv8m. To further improve detection accuracy, a more lightweight dynamic upsampling operator, DySample, is utilized in the feature fusion network to enhance the quality and richness of the fused features. The YOLOv8-A model is evaluated on an industrially provided dataset, and experimental results indicate that, compared to the original algorithm, the parameter count and computational load of YOLOv8-A are reduced to 2.19 × 10^7, while achieving a 98.2% mAP and improving small target detection by 5.8 percentage points. The effectiveness and superiority of this method are validated through comparisons with other mainstream target detection algorithms.