基于AI视觉识别的薄壁模具表面缺陷检测数据

通过工业相机采集薄壁模具表面图像，利用深度学习算法识别划痕、凹陷、毛刺等7类典型缺陷的技术，融合YOLOv5目标检测与ResNet34分类网络，实现对塑料模具表面多维缺陷的自动化识别。本缺陷检测数据具有以下应用场景：在企业内部，1. 通过实时采集模具表面缺陷数据，构建注塑工艺参数与缺陷类型的关联模型，实现工艺参数动态调整。2.可持续监测模具缺陷演变规律，建立磨损生命周期预测模型。在企业外部，1. 积累的数万条缺陷样本数据库，为行业协会制定《塑料模具表面缺陷分级标准》提供数据支撑。2. 与机器视觉厂商合作开发专用检测设备，集成本缺陷识别模型后，使同类企业检测效率大大提升。首先，图像采集：采用2000万像素工业相机，配合环形LED光源多角度照射。其次，特征提取：通过CNN卷积层提取缺陷纹理、几何等特征。再次，量化模型，构建缺陷指数公式：P=w1*S+w2*D+w3*C+w4*T+w5*L，其中，w1至w5为各个输入量参数的权重，其大小值利用Halcon深度学习框架构建缺陷检测模型时，通过反向传播算法自动调整大小。S为缺陷面积比，D为最大深度，C为颜色对比度差异（相邻阶灰度差≥5%时C为0.25，相邻阶灰度差≥10%时C为0.45，相邻阶灰度差≥15%时C为0.67，否则C为0），T为缺陷类型系数（划痕T=0.6、毛刺T=0.55、裂纹T=0.85，否则T为0），L为缺陷位置影响因子（边缘L=0.7，中心L=0.55，否则L为0）。最后，根据计算得到的P值对薄壁模具表面缺陷进行分级，优级(P≤0.25)，良级(0.25<P≤0.48)，差级(P>0.48)。

This technology collects surface images of thin-walled molds via industrial cameras, uses deep learning algorithms to identify 7 typical defect types including scratches, dents, burrs, etc., and fuses YOLOv5 object detection and ResNet34 classification networks to realize automatic recognition of multi-dimensional defects on plastic mold surfaces.

This defect detection data has the following application scenarios:
1. Internal enterprise scenarios:
- Collect mold surface defect data in real time to build a correlation model between injection molding process parameters and defect types, and achieve dynamic adjustment of process parameters.
- Continuously monitor the evolution law of mold defects and establish a wear life cycle prediction model.
2. External enterprise scenarios:
- The accumulated tens of thousands of defect sample databases provide data support for industry associations to formulate the "Classification Standards for Surface Defects of Plastic Molds".
- Cooperate with machine vision manufacturers to develop specialized testing equipment; after integrating this defect recognition model, the detection efficiency of peer enterprises will be greatly improved.

The specific implementation steps are as follows:
First, Image Collection: Adopt a 20-megapixel industrial camera, paired with a ring-shaped LED light source for multi-angle illumination.
Second, Feature Extraction: Extract defect texture, geometric and other features through CNN convolutional layers.
Third, Quantitative Model Construction: Establish the defect index formula: P=w1*S+w2*D+w3*C+w4*T+w5*L, where w1 to w5 are the weights of each input parameter, which are automatically adjusted via the backpropagation algorithm when building a defect detection model using the Halcon deep learning framework. The specific meanings of each parameter are:
- S: Defect area ratio
- D: Maximum defect depth
- C: Color contrast difference (C=0.25 when the adjacent grayscale difference ≥5%, C=0.45 when ≥10%, C=0.67 when ≥15%, otherwise C=0)
- T: Defect type coefficient (T=0.6 for scratch, T=0.55 for burr, T=0.85 for crack, otherwise T=0)
- L: Defect location impact factor (L=0.7 for edge, L=0.55 for center, otherwise L=0)
Finally, Defect Grading: Classify the surface defects of thin-walled molds based on the calculated P value: excellent grade (P ≤ 0.25), good grade (0.25 < P ≤ 0.48), and poor grade (P > 0.48).

搜集汇总

背景与挑战

背景概述

该数据集是一个基于AI视觉识别的薄壁模具表面缺陷检测数据集，包含10237条记录，每日更新，数据结构详细，涵盖21个字段，主要用于企业内部工艺优化和外部行业标准制定及设备开发。