Google Vision OCR无法识别小文本及倒置图像问题咨询

Hey there, let's fix those two common Google Vision OCR problems you're running into—here are practical, code-ready solutions tailored to your Camera2 + Vision setup:

1. Fixing Inverted Image Recognition #

Camera2 captures images based on the sensor's native orientation, which often doesn't match your device's display orientation. This causes the OCR to process upside-down or rotated text, leading to failed recognition.

How to Correct Image Orientation: #

• Calculate the required rotation angle by matching the device's display rotation to the camera sensor's orientation.
• Rotate the captured bitmap before passing it to the Frame object.

Code Modification: #

First, add a helper method to get the correct rotation angle:

private int getCorrectRotationAngle(CameraCharacteristics characteristics, int displayRotation) {
    int sensorOrientation = characteristics.get(CameraCharacteristics.SENSOR_ORIENTATION);
    int rotationCompensation;

    switch (displayRotation) {
        case Surface.ROTATION_0:
            rotationCompensation = sensorOrientation;
            break;
        case Surface.ROTATION_90:
            rotationCompensation = (sensorOrientation + 270) % 360;
            break;
        case Surface.ROTATION_180:
            rotationCompensation = (sensorOrientation + 180) % 360;
            break;
        case Surface.ROTATION_270:
            rotationCompensation = (sensorOrientation + 90) % 360;
            break;
        default:
            rotationCompensation = sensorOrientation;
            break;
    }
    return rotationCompensation;
}

Then rotate your bitmap before creating the Frame:

// Get the rotation angle (pass CameraCharacteristics from your Camera2 setup)
int rotationAngle = getCorrectRotationAngle(cameraCharacteristics, getWindowManager().getDefaultDisplay().getRotation());

// Rotate the bitmap
Matrix matrix = new Matrix();
matrix.postRotate(rotationAngle);
Bitmap rotatedBitmap = Bitmap.createBitmap(bitmap, 0, 0, bitmap.getWidth(), bitmap.getHeight(), matrix, true);

// Now use the rotated bitmap for OCR
Frame frame = new Frame.Builder().setBitmap(rotatedBitmap).build();
SparseArray<TextBlock> detectedItems = textRecognizer.detect(frame);

2. Improving Small Text Recognition #

Google Vision OCR performs better with high-resolution, clear text. For small text, we can boost recognition by preprocessing the image and adjusting OCR settings.

Key Fixes: #

• Scale Up the Image: Enlarge the bitmap (focus on the text region if possible) to increase text size relative to the OCR's detection threshold.
• Image Preprocessing: Convert to grayscale or apply contrast enhancement to make small text more distinct.
• Target ROI (Region of Interest): If you know where the small text is, crop that area and process it separately to avoid wasting resources on irrelevant parts.

Code Examples: #

Scaling the Bitmap:

// Scale the bitmap by 2x (adjust factor based on your needs)
float scaleFactor = 2.0f;
Bitmap scaledBitmap = Bitmap.createScaledBitmap(bitmap, 
    (int)(bitmap.getWidth() * scaleFactor), 
    (int)(bitmap.getHeight() * scaleFactor), 
    true);

// Use scaled bitmap for OCR
Frame frame = new Frame.Builder().setBitmap(scaledBitmap).build();

Preprocessing for Contrast:

// Convert bitmap to grayscale and enhance contrast
Bitmap grayscaleBitmap = Bitmap.createBitmap(bitmap.getWidth(), bitmap.getHeight(), Bitmap.Config.ARGB_8888);
Canvas canvas = new Canvas(grayscaleBitmap);
Paint paint = new Paint();
ColorMatrix colorMatrix = new ColorMatrix();
colorMatrix.setSaturation(0); // Grayscale
ColorMatrixColorFilter filter = new ColorMatrixColorFilter(colorMatrix);
paint.setColorFilter(filter);
canvas.drawBitmap(bitmap, 0, 0, paint);

// Enhance contrast (adjust values as needed)
ColorMatrix contrastMatrix = new ColorMatrix();
contrastMatrix.set(new float[]{
    1.5f, 0, 0, 0, -50, // Red channel
    0, 1.5f, 0, 0, -50, // Green channel
    0, 0, 1.5f, 0, -50, // Blue channel
    0, 0, 0, 1, 0      // Alpha channel
});
paint.setColorFilter(new ColorMatrixColorFilter(contrastMatrix));
Canvas contrastCanvas = new Canvas(grayscaleBitmap);
contrastCanvas.drawBitmap(grayscaleBitmap, 0, 0, paint);

// Use preprocessed bitmap for OCR
Frame frame = new Frame.Builder().setBitmap(grayscaleBitmap).build();

Combining Both Fixes in Your Existing Code:

Here's how you can integrate both fixes into your original code:

TextRecognizer textRecognizer = TextRecognizerController.getControllerInstance().getTextRecognizerInstance(getContext(), getView()); 
if (textRecognizer != null && textRecognizer.isOperational()) { 
    // Step 1: Correct image orientation
    int rotationAngle = getCorrectRotationAngle(cameraCharacteristics, getWindowManager().getDefaultDisplay().getRotation());
    Matrix matrix = new Matrix();
    matrix.postRotate(rotationAngle);
    Bitmap rotatedBitmap = Bitmap.createBitmap(bitmap, 0, 0, bitmap.getWidth(), bitmap.getHeight(), matrix, true);

    // Step 2: Scale and preprocess for small text
    float scaleFactor = 2.0f;
    Bitmap scaledBitmap = Bitmap.createScaledBitmap(rotatedBitmap, 
        (int)(rotatedBitmap.getWidth() * scaleFactor), 
        (int)(rotatedBitmap.getHeight() * scaleFactor), 
        true);

    // Create frame with processed bitmap
    Frame frame = new Frame.Builder().setBitmap(scaledBitmap).build(); 
    SparseArray<TextBlock> detectedItems = textRecognizer.detect(frame); 
    if (detectedItems != null && detectedItems.size() != 0) {
        // Your existing processing logic here
    }
}

Additional Tips: #

• Ensure your TextRecognizer is using the latest model by enabling auto-update in the Vision API settings.
• If possible, guide users to align small text within the center of the camera frame for better capture quality.

内容的提问来源于stack exchange，提问作者Ragini