当前位置:网站首页>Practical example of propeller easydl: automatic scratch recognition of industrial parts
Practical example of propeller easydl: automatic scratch recognition of industrial parts
2022-07-05 16:56:00 【Paddlepaddle】
Baidu PaddlePaddle EasyDL Provide zero threshold for enterprise developers AI Development platform , One stop support for smart tagging 、 model training 、 Service deployment and other functions , Built in rich pre training model , Support for public clouds / Local server / Flexible deployment on the device side , It has been used in industry 、 retail 、 manufacture 、 Medical treatment and other fields .
Automatic scratch recognition of industrial parts
Through training object detection customization AI Model , To achieve for “ Scratches on industrial parts ” Intelligent recognition of , It is widely used in industrial quality inspection 、 Intelligent manufacturing and other fields .
Project description
Business background
A factory is mainly engaged in the production and processing of various workpieces , Because the factory production environment is complex , There are scratches in the production process of the workpiece 、 Bump 、 Possibility of making omissions , Defective goods should be treated as defective products , Do not enter the market .
Business difficulties
The precision of defect quality inspection is required to be high , Defects are usually subtle and difficult to find , The traditional manual quality inspection method is inefficient , Long time operation is prone to visual fatigue and false detection 、 Phenomenon of missing inspection . Low quality inspection efficiency will directly affect the factory's production capacity and delivery efficiency .
Solutions
To improve production capacity and quality inspection accuracy , The factory decided to intelligently empower the quality inspection production line , adopt AI Realize automatic identification of defects + Manual review method , Improve the efficiency of ex factory quality inspection . Learn about the oars EasyDL Provides zero threshold AI Development function , Model training can be completed quickly without knowing the basis of the algorithm , The factory passed EasyDL The object detection task type provided by the platform is quickly trained in the following steps , High precision scratch quality inspection of production workpiece is obtained AI Model and put it into practical business application , Based on this intelligent quality inspection scheme , The quality inspection manpower invested by the enterprise every year is reduced 60%, The efficiency and accuracy of quality inspection have been greatly improved .
Tips : For workpiece surface scratch recognition scene , Multiple scratch defects may appear in a single picture , The number of scratches needs to be identified , Therefore, the object detection model is selected .
Data preparation
Data collection and import
The factory selects historical defect data from the original production environment, about 300 Zhang , When preparing training data, we should pay attention to keeping consistent with the data generated by the actual production scene as far as possible , Like light 、 brightness 、 Shooting angle, etc , Only when the objective factors are consistent, the trained model will have higher applicability and effect in practical application . for example , The angle of the camera deployed in the actual quality inspection production line to the parts is overhead , that , When collecting data, you also need to collect the original pictures from the same angle . If the original data obtained by the enterprise has no annotation information , You can choose to import unmarked information in the platform and then press one key to import , for example , You can type the original image into a compressed package , The specific import path is as follows :
meanwhile EasyDL It also provides a variety of other data import methods , You can choose different import methods according to your usage habits , for example , You can upload the picture to the network address first ( for example , Baidu SkyDrive 、 Baidu BOS etc. ), Copy the corresponding address link to complete the quick import , It is suitable for cases with large data magnitude , Help you shorten the import time .
Data tagging
The original image needs to be marked , Before model training . The data marking process should be as close to the defect characteristics of scratches as possible , Make sure that all defects are contained in the rectangular box and that the rectangular box is not too large than the defect size , The following are examples of correct annotation and wrong annotation .
Tips : When your annotation workload is large , You can choose to turn on smart annotation , That is, you only need to mark 30% Pictures of the , The remaining 70% The picture of will automatically label you .
model training
After training data preparation , You can create it by clicking 【 Creating models 】 Enter the basic information and complete the model creation , Click on 【 Training models 】 Start training . A model can cover multiple versions , One training produces one version , You can customize the version management according to your business needs . Considering that industrial quality inspection scenarios usually pay more attention to prediction efficiency , The prediction efficiency of the model will directly affect the quality inspection efficiency , Therefore, it is recommended that you choose EasyEdge Local deployment - The server - Ultra high precision algorithm . The advantages of this scheme are : The local deployment method can effectively save the time required for the upload of prediction data and the distribution of prediction results , Greatly improve the prediction efficiency and then increase the production capacity , Ultra high precision algorithm can maximize the effect of the model , Here's the picture :
Model deployment
Local deployment requires local computing power to carry the model deployment and prediction process , To ensure the prediction performance of the model , The factory chose the one with sufficient computing power T4 Server as local computing power , Therefore, it is necessary to publish the model as a model suitable for the hardware SDK. meanwhile EasyDL The platform still supports many other hardware types , You can choose the corresponding operating system and chip type according to your actual business needs , Here's the picture :
After the model is released , Can be found in 【 Pure offline service 】 Found published models in SDK, Download and 【 Get the serial number 】 After activation, it can be deployed in the terminal application .
The enterprise will AI The model is applied to the actual scene, hoping to realize the automatic sorting of defective workpieces , In the quality inspection production line, each workpiece is photographed by a professional camera , The captured pictures are transferred to the prediction server for quality inspection , When defects are detected in the current workpiece picture , The defective workpiece is grabbed by the mechanical arm and stored in a specific area for manual review .
Effect optimization
After model training , adopt EasyDL The overall effect of the model can be viewed in the automatically produced evaluation report , In this factory scenario, the defect characteristics are relatively consistent and the scenario is fixed , The overall effect is better . If the effect is not good, you can check the error examples of different labels to locate the reasons for the poor performance under this label . for example , By looking at the error examples , Pictures with wrong predictions are usually taken during periods of insufficient light , For optimization suggestions, we need to supplement the data in the case of insufficient lighting Improve model effect , Here's the picture :
common problem
problem 1:EasyDL Images provide a variety of training tasks , Why choose the object detection task type in this scenario ?
at present EasyDL Images provide image classification 、 Object detection 、 There are three types of tasks in image segmentation : Image classification is applicable to scenes with a single subject or state in the picture ; Object detection is applicable to scenes with multiple subjects and the need to identify the location of the subjects ; Image segmentation can locate multiple subjects and needs to be located at the pixel level of the subject . For the scratch inspection scenario , Multiple defects may appear in a single picture , Defects do not need to be located at the pixel level , Therefore, it is more appropriate to select object detection .
problem 2: How to collect industrial quality inspection scene data ?
The data is AI Prerequisites for model training , The training data in the industrial quality inspection scene usually comes from the historical stock data of the image acquisition equipment on the industrial production line , therefore , If image acquisition equipment has not been built on the production line, priority should be given to the construction of image acquisition equipment , Based on the collected image data AI model training .
problem 3: How to solve the problem of too few defect samples ? EasyDL The platform provides data enhancement strategies , in the light of EasyDL The platform provides data imbalance strategies , The existing defect pictures can be expanded by data enhancement strategy , And then maximize the utilization of limited data , Complete the configuration on your model training page .
边栏推荐
- 深潜Kotlin协程(二十一):Flow 生命周期函数
- 【剑指 Offer】66. 构建乘积数组
- [es6] add if judgment or ternary operator judgment in the template string
- 數據訪問 - EntityFramework集成
- Writing method of twig array merging
- PHP talent recruitment system development source code recruitment website source code secondary development
- 如何安装mysql
- 精准防疫有“利器”| 芯讯通助力数字哨兵护航复市
- Google Earth engine (GEE) -- a brief introduction to kernel kernel functions and gray level co-occurrence matrix
- 拷贝方式之DMA
猜你喜欢
深潜Kotlin协程(二十一):Flow 生命周期函数
Scratch colorful candied haws Electronic Society graphical programming scratch grade examination level 3 true questions and answers analysis June 2022
2020-2022两周年创作纪念日
Learnopongl notes (II) - Lighting
机器学习编译第2讲:张量程序抽象
Enter a command with the keyboard
Flet tutorial 12 stack overlapping to build a basic introduction to graphic and text mixing (tutorial includes source code)
Deep learning plus
2020-2022 two-year anniversary of creation
文件操作--I/O
随机推荐
[js] skill simplification if empty judgment
Explain in detail the functions and underlying implementation logic of the groups sets statement in SQL
Apple has abandoned navigationview and used navigationstack and navigationsplitview to implement swiftui navigation
关于new Map( )还有哪些是你不知道的
scratch五彩糖葫芦 电子学会图形化编程scratch等级考试三级真题和答案解析2022年6月
[team PK competition] the task of this week has been opened | question answering challenge to consolidate the knowledge of commodity details
Solve cmakelist find_ Package cannot find Qt5, ECM cannot be found
Enter a command with the keyboard
【机器人坐标系第一讲】
阈值同态加密在隐私计算中的应用:解读
Summary of methods for finding intersection of ordered linked list sets
搜索 正排索引 和 倒排索引 区别
Yarn common commands
Get ready for the pre-season card game MotoGP ignition champions!
Deep learning plus
File operation --i/o
Dare not buy thinking
Win11 prompt: what if the software cannot be downloaded safely? Win11 cannot download software safely
Google Earth engine (GEE) -- a brief introduction to kernel kernel functions and gray level co-occurrence matrix
Apple 已弃用 NavigationView,使用 NavigationStack 和 NavigationSplitView 实现 SwiftUI 导航