Spring Boot + Tess4j: Build OCR REST API Step by Step

Introduction

Java can implement OCR (Optical Character Recognition) to automatically extract text from images using Spring Boot and Tess4j, a Java wrapper for the Tesseract engine. This enables a lightweight RESTful service for various OCR needs without bulky external tools.

OCR Engine Comparison

Four popular OCR engines are compared:

Tesseract (Google) : 100+ languages, average Chinese optimization, good accuracy, medium speed, medium model size, simple Spring Boot integration, high maintenance activity.

PaddleOCR (Baidu) : 80+ languages, excellent Chinese optimization, excellent accuracy, fast speed, larger model size, medium Spring Boot integration, very high maintenance activity.

EasyOCR (aided AI) : 80+ languages, good Chinese optimization, good accuracy, fast speed, medium model size, simple Spring Boot integration, high maintenance activity.

TrOCR (Microsoft) : Multi-language, good Chinese optimization, excellent accuracy, medium speed, medium model size, complex Spring Boot integration, medium maintenance activity.

Function Demonstration

Result of OCR detection on a sample image:

Implementation

1. Description

Spring Boot combined with Tess4j provides a Java-based OCR solution. Tess4j wraps Tesseract, allowing easy integration into Java applications. The stack yields a lightweight RESTful service for OCR tasks.

2. Coding Implementation

2.1 Add Dependency

<dependency>
<groupId>net.sourceforge.tess4j</groupId>
<artifactId>tess4j</artifactId>
</dependency>

2.2 Initialize Tesseract Engine

Deployment considerations:

Using

new ClassPathResource("tess_data").getFile().getAbsolutePath()

may fail after packaging as a JAR. Refer to open-source projects like TensorflowUtil to copy resource files before loading.

In Linux, resolve net.sourceforge.tess4j.TessAPI initialization issues by ensuring all native dependencies and system configurations are correct.

Training data: Different training data affects accuracy and speed. Free training data sets include: tessdata_best: High accuracy, slower recognition. tessdata: Standard balance of speed and accuracy. tessdata_fast: Faster recognition, slightly lower accuracy.

/** TesseractOcr model loading */
@Slf4j
@Getter
@Component
public class TesseractOcrModelService {
    private final Tesseract tesseract = new Tesseract();

    public TesseractOcrModelService() {
        try {
            // Get training model folder (this method has issues when packaged as JAR; use TensorflowUtil)
            String folderPath = new ClassPathResource("tess_data").getFile().getAbsolutePath();
            /* OEM modes:
             * OEM_TESSERACT_ONLY = 0: Tesseract only
             * OEM_LSTM_ONLY = 1: LSTM only
             * OEM_TESSERACT_LSTM_COMBINED = 2: Combined
             * OEM_DEFAULT = 3: Auto
             */
            tesseract.setOcrEngineMode(2); // OEM_TESSERACT_LSTM_COMBINED
            // Set training data path
            tesseract.setDatapath(folderPath);
            tesseract.setPageSegMode(6);
            // Set language to Simplified Chinese
            tesseract.setLanguage("chi_sim");
        } catch (Exception e) {
            throw new RuntimeException(e);
        }
    }
}

2.3 Write RESTful Interface

/** OCR Controller */
@RestController
@RequestMapping("/ocr")
@RequiredArgsConstructor
public class OcrController {
    private final TesseractOcrModelService tesseractOcrModelService;

    @PostMapping("/detection")
    public Result<String> ocrDetection(MultipartFile file) {
        try {
            /* Image preprocessing recommendations:
             * Binarization: convert to black/white for contrast
             * Denoising: remove noise
             * Rotation correction: ensure text is horizontal
             */
            Tesseract tesseract = tesseractOcrModelService.getTesseract();
            return Result.success(tesseract.doOCR(ImageIO.read(file.getInputStream())));
        } catch (Exception e) {
            throw new RuntimeException("ImageIO.read(file.getInputStream()) parsing error");
        }
    }
}

Source Code

https://gitee.com/fateyifei/yf

Conclusion

Tess4j performs well for ID numbers, phone numbers, and English words, but Chinese recognition with free training data is relatively poor. For higher quality requirements, consider:

Custom training: Train with custom datasets to improve accuracy for specific text types or languages.

Third-party APIs: Use professional OCR services like Google Cloud Vision, Microsoft Azure OCR, or Amazon Textract for higher accuracy and more features.

Additional applications of Tess4j include:

Document digitization: Convert paper documents to editable electronic text.

Automated data entry: Extract data from scanned forms, invoices, etc.

License plate recognition: Identify plate numbers from traffic camera images.

Handwriting recognition: Convert handwritten content to digital text.

Although these approaches may involve extra cost and setup, they significantly improve recognition results to meet higher demands.

Reference

juejin.cn/post/7408619644195766272