Accelerating Quality Control: Leveraging OCR for Automated Batch Code Inspection

Speeding Up Quality Control: Automated Batch Code Reading

Introduction to OCR

OCR is a technology that extracts text from scanned paper documents, PDF files, images, and videos captured by a digital camera. It converts visual representations of text into machine-encoded digital text that can be edited, searched, and processed. This extracted text is further used to conduct data analysis enabling automation and improved accessibility.

Use Cases of OCR in Industrial Automation

OCR is commonly used in document processing for tasks such as invoice processing, document digitization, data extraction from forms, automating data entry, and enhancing document searchability. While OCR has been widely adopted for tasks such as digitizing printed documents, it also has crucial applications in industrial automation and quality control. One of the most common uses of OCR technology in manufacturing plants is batch code inspection. This system ensures product quality, traceability, and consumer safety across manufacturing industries. It accurately and efficiently extracts batch codes, expiry dates, serial numbers, and other critical information from product packaging. Such information is essential for tracking and verifying product authenticity, managing inventory, complying with regulatory requirements, and swiftly addressing product recalls in case of defects or contamination.

Challenges in OCR Implementation

Using OCR for batch code inspection in a high-speed manufacturing plant presents significant challenges due to diverse fonts, formats, and printing conditions. Generic OCR engines excel with standard fonts and optimal lighting but struggle with non-standard fonts, poor lighting, or difficult surfaces. To achieve high accuracy quickly, it’s crucial to employ a trainable, lightweight model tailored to these specific conditions, ensuring rapid and precise inspection even under demanding manufacturing scenarios.

OCR Solutions for Batch Code Inspection

PaddleOCR is an excellent solution for addressing the challenges of batch code inspection. It’s a multilingual OCR toolkit that provides practical tools and high-quality pre-trained models for text detection, direction classification, and recognition. With just a few lines of code, users can apply and train models, achieving accuracy comparable to commercial OCR products.

Flow Chart of Batch Code Inspection System

Data Augmentation

It involves two steps i.e. Text Detection and Text Recognition and it is important to train the text detection and recognition model with a variety of data to increase the robustness of the model. However, it is difficult to collect all the possible variations of data in real-time, hence data augmentation techniques will be used to generate data with many variations that mimic industrial real-time data.

Detection

Every image is augmented with various augmentation techniques recursively and randomly. Augmentation techniques such as rotation, blur, contrast adjustment, flipping, and elastic and rigid transformation are applied. For customization in augmentation, each technique is assigned a probability or weight value.

Recognition

With Data Augmentation techniques, variations of data are generated for text detection. However, for text recognition variations in character numbers and symbols are necessary. Data Fabrication techniques are used to accomplish this. For this, a blank image is used as the canvas, and the text fonts are loaded. Random characters are generated from these loaded fonts in the batch code format. By randomly arranging the batch code positions from horizontal and vertical and applying random rotation angles, this technique generates a high degree of randomization. As a result, a variety of data are used to make both text detection and recognition models more generalized and robust.

Text Detection

For text detection, DBNet is used which is a text detection algorithm that uses Differentiable Binarization to detect text in complex scenes, including curved or rotated text. Instead of using a fixed threshold to separate text from the background, DBNet learns the best way to binarize the image during training. This makes it highly accurate and adaptable to various lighting and text shapes. It’s also designed for real-time applications, making it fast and reliable for detecting text in both images and videos. It consists of :

• Backbone: A feature extractor, often based on a pre-trained CNN like ResNet, that outputs multi-level feature maps.
• Feature Fusion: A feature pyramid is used to merge features across different scales to capture text at various sizes.
• Probability Map: A differentiable binarization map predicts text areas using adaptive thresholding.
• Text Region Segmentation: It generates precise text boundaries, making DBNet highly effective for irregular and curved text shapes.

Text Recognition

For text recognition, SVTR is used which is a powerful and flexible model for text recognition, designed to handle complex and irregular text structures with high accuracy. It uses spatial attention and a transformer-based architecture to capture relationships between characters, making it highly accurate in recognizing text in challenging conditions. By focusing on both character details and overall structure, SVTR delivers robust performance for real-world text recognition tasks. It consists of:

• Progressive Overlapping Patch Embedding: This module processes the input image by splitting it into patches through two consecutive convolution layers with strides of 2. This technique progressively increases the feature dimensions while maintaining better feature fusion compared to single-step projections, which is essential for capturing text-specific details
• Mixing Block: The model employs two types of mixing blocks:
• Global Mixing: This block uses multi-head self-attention to model long-term dependencies between different text and non-text components, helping to reduce noise from non-text areas.
• Local Mixing: Here, the model focuses on short-range dependencies within a predefined window, capturing fine-grained character features, such as strokes and morphology
• Merging: After each mixing block stage (except the last one), a merging operation is applied. This reduces the height dimension of the feature map while preserving the width. This is particularly useful for scene text, which is mostly horizontally oriented
• Combining and Prediction: In the final stage, a “combining” operation reduces the height dimension to 1, converting the feature map into a sequence suitable for recognition. The model then uses a linear classifier to predict characters from this sequence

Conclusion

OCR technology plays a vital role in industrial automation and quality control, enabling efficient data extraction, analysis, and automation of critical processes. In batch code inspection, OCR ensures product traceability, authenticity, and regulatory compliance by accurately extracting essential information like batch codes, expiry dates, and serial numbers. However, implementing OCR in high-speed manufacturing environments presents challenges due to varying fonts, surfaces, and lighting conditions. Solutions like PaddleOCR, coupled with augmentation techniques for text detection and recognition, enhance model robustness and accuracy, making them well-suited for industrial applications. By leveraging advanced text detection algorithms like DBNet and text recognition models such as SVTR, manufacturers can achieve high-performance OCR systems capable of handling complex and irregular text. These advancements in OCR technology ensure better quality control, consumer safety, and efficient operations in modern manufacturing environments.

#OCR | #ComputerVision | #AI | #MachineLearning | #QualityControl | #IndustrialAutomation | #Manufacturing | #BatchCode | #ProductTraceability | #DataDriven | #Technology | #Technology | #paddle | #industrialAutomation | #industry4.0 | #ML

Link to Author's LinkedIn profile: