Output image is weired while trying to inference esrgan, Someone help me please.

[md-rifatkhan]

I'm trying to inference real esrgan, but cant able to get output correctly. I'm using com.microsoft.onnxruntime:onnxruntime-android:1.17.3

Model Link: Google Drive

Inference Class:

public class ImageInference {

    private static final String TAG = "ImageInference";

    // Load the image from assets
    public static Bitmap loadImageFromAssets(Context context, String fileName) throws IOException {
        Log.d(TAG, "Loading image from assets: " + fileName);
        InputStream is = context.getAssets().open(fileName);
        Bitmap image = BitmapFactory.decodeStream(is);
        if (image != null) {
            Log.d(TAG, "Image loaded successfully: " + fileName);
        } else {
            Log.d(TAG, "Failed to load image: " + fileName);
        }
        return image;
    }

    // Convert Bitmap to FloatBuffer for ONNX Runtime
    public static FloatBuffer bitmapToFloatBuffer(Bitmap bitmap, float mean, float std) {
        int width = bitmap.getWidth();
        int height = bitmap.getHeight();
        Log.d(TAG, "Preparing to convert bitmap to FloatBuffer. Width: " + width + ", Height: " + height);

        int[] pixels = new int[width * height];
        bitmap.getPixels(pixels, 0, width, 0, 0, width, height);
        FloatBuffer buffer = FloatBuffer.allocate(width * height * 3);

        for (final int val : pixels) {
            buffer.put(((val >> 16) & 0xFF) / 255.f - mean / std); // RED
            buffer.put(((val >> 8) & 0xFF) / 255.f - mean / std);  // GREEN
            buffer.put((val & 0xFF) / 255.f - mean / std);         // BLUE
        }
        buffer.flip(); // Prepare buffer for reading
        Log.d(TAG, "Bitmap successfully converted to FloatBuffer.");
        return buffer;
    }

    public static Bitmap tensorToBitmap(float[][][][] tensor) {
        // Assume tensor dimensions are [1][3][height][width]
        int height = tensor[0][0].length;
        int width = tensor[0][0][0].length;

        Bitmap bitmap = Bitmap.createBitmap(width, height, Bitmap.Config.ARGB_8888);
        int[] pixels = new int[width * height];

        for (int y = 0; y < height; y++) {
            for (int x = 0; x < width; x++) {
                int r = (int) (tensor[0][0][y][x] * 255);
                int g = (int) (tensor[0][1][y][x] * 255);
                int b = (int) (tensor[0][2][y][x] * 255);
                pixels[y * width + x] = 0xFF000000 | (r << 16) | (g << 8) | b;
            }
        }

        bitmap.setPixels(pixels, 0, width, 0, 0, width, height);
        return bitmap;
    }

    public static float[][][][] runInference(Context context, String modelPath, Bitmap image) throws OrtException, IOException {

        InputStream is = context.getAssets().open(modelPath);
        ByteArrayOutputStream bos = new ByteArrayOutputStream();
        byte[] buf = new byte[1024];
        for (int readNum; (readNum = is.read(buf)) != -1;) {
            bos.write(buf, 0, readNum);
        }
        byte[] modelBytes = bos.toByteArray();

        Log.d(TAG, "Starting inference with model: " + modelPath);

        OrtEnvironment env = OrtEnvironment.getEnvironment();
        InputStream modelInputStream = context.getAssets().open(modelPath);
        OrtSession session = env.createSession(modelBytes, new OrtSession.SessionOptions());
        Log.d(TAG, "Model and Session created successfully." );
        try {
            int width = image.getWidth();
            int height = image.getHeight();
            FloatBuffer inputBuffer = bitmapToFloatBuffer(image, 0f, 1f);
            OnnxTensor tensor = OnnxTensor.createTensor(env, inputBuffer, new long[]{1, 3, height, width});
            OrtSession.Result results = session.run(Collections.singletonMap("input", tensor));
            float[][][][] output = (float[][][][]) results.get(0).getValue();
            tensor.close();
            Log.d(TAG, "Inference completed successfully.");
            return output;
        } finally {
            session.close();
            modelInputStream.close();
            env.close();
            Log.d(TAG, "Cleaned up ONNX resources.");
        }
    }
}

Main Activity

public class MainActivity extends Activity {

    private ImageView orginalImageView;
    private ImageView outputImageView;
    private TextView textViewResult;

    @Override
    protected void onCreate(Bundle savedInstanceState) {
        super.onCreate(savedInstanceState);
        setContentView(R.layout.activity_main);

        orginalImageView = findViewById(R.id.orginalImage);
        outputImageView = findViewById(R.id.outputImage);
        textViewResult = findViewById(R.id.textViewResult);

        try {
            Bitmap image = ImageInference.loadImageFromAssets(this, "LR.png");
            orginalImageView.setImageBitmap(image);
            float[][][][] output = ImageInference.runInference(this, "realesr-general-x4v3-fp32.onnx", image);
            if (output != null && output.length > 0 && output[0].length > 0 && output[0][0].length > 0) {
                int height = output[0][0].length;
                int width = output[0][0][0].length;

                Bitmap outputImage = ImageInference.tensorToBitmap(output);
                outputImageView.setImageBitmap(outputImage);  // Display the output image
                textViewResult.setText("Inference complete with output shape: " + "Height " + height + " Width "  + width);
            } else {
                textViewResult.setText("Inference complete but no valid output!");
            }
        } catch (Exception e) {
            textViewResult.setText("Inference failed: " + e.getMessage());
            e.printStackTrace();
        }
    }

[md-rifatkhan]