nv-ingest-api 2025.7.13.dev20250713__py3-none-any.whl → 2025.7.15.dev20250715__py3-none-any.whl

nv_ingest_api/internal/extract/pdf/engines/nemoretriever.py

@@ -40,6 +40,7 @@ from nv_ingest_api.internal.schemas.meta.metadata_schema import validate_metadat
 from nv_ingest_api.internal.primitives.nim.model_interface.yolox import (
     YOLOX_PAGE_IMAGE_PREPROC_WIDTH,
     YOLOX_PAGE_IMAGE_PREPROC_HEIGHT,
+    YOLOX_PAGE_IMAGE_FORMAT,
 )
 from nv_ingest_api.internal.schemas.extract.extract_pdf_schema import NemoRetrieverParseConfigSchema
 from nv_ingest_api.util.metadata.aggregators import (
@@ -355,7 +356,7 @@ def nemoretriever_parse_extractor(
                 img_numpy = crop_image(page_image, transformed_bbox)
 
                 if img_numpy is not None:
-                    base64_img = numpy_to_base64(img_numpy)
+                    base64_img = numpy_to_base64(img_numpy, format=YOLOX_PAGE_IMAGE_FORMAT)
                     image = Base64Image(
                         image=base64_img,
                         bbox=transformed_bbox,

nv_ingest_api/internal/extract/pdf/engines/pdfium.py

@@ -28,6 +28,7 @@ from nv_ingest_api.internal.primitives.nim.default_values import YOLOX_MAX_BATCH
 from nv_ingest_api.internal.primitives.nim.model_interface.yolox import (
     YOLOX_PAGE_IMAGE_PREPROC_WIDTH,
     YOLOX_PAGE_IMAGE_PREPROC_HEIGHT,
+    YOLOX_PAGE_IMAGE_FORMAT,
     get_yolox_model_name,
     YoloxPageElementsModelInterface,
 )
@@ -186,7 +187,7 @@ def _extract_page_element_images(
             if cropped is None:
                 continue
 
-            base64_img = numpy_to_base64(cropped)
+            base64_img = numpy_to_base64(cropped, format=YOLOX_PAGE_IMAGE_FORMAT)
 
             bbox_in_orig_coord = (
                 int(w1) - pad_width,

nv_ingest_api/internal/primitives/nim/model_interface/nemoretriever_parse.py

@@ -120,6 +120,7 @@ class NemoRetrieverParseModelInterface(ModelInterface):
             logger.debug("Formatting input for HTTP NemoRetrieverParse model")
             # Prepare payload for HTTP request
 
+            ## TODO: Ask @Edward Kim if we want to switch to JPEG/PNG here
             if "images" in data:
                 base64_list = [numpy_to_base64(img) for img in data["images"]]
             else:

nv_ingest_api/internal/primitives/nim/model_interface/yolox.py

@@ -2,9 +2,7 @@
 # All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-
-import base64
-import io
+import os
 import logging
 import warnings
 from math import log
@@ -20,11 +18,11 @@ import packaging
 import pandas as pd
 import torch
 import torchvision
-from PIL import Image
 
 from nv_ingest_api.internal.primitives.nim import ModelInterface
 from nv_ingest_api.internal.primitives.nim.model_interface.helpers import get_model_name
 from nv_ingest_api.util.image_processing import scale_image_to_encoding_size
+from nv_ingest_api.util.image_processing.transforms import numpy_to_base64
 
 logger = logging.getLogger(__name__)
 
@@ -35,6 +33,7 @@ YOLOX_PAGE_MIN_SCORE = 0.1
 YOLOX_PAGE_NIM_MAX_IMAGE_SIZE = 512_000
 YOLOX_PAGE_IMAGE_PREPROC_HEIGHT = 1024
 YOLOX_PAGE_IMAGE_PREPROC_WIDTH = 1024
+YOLOX_PAGE_IMAGE_FORMAT = os.getenv("YOLOX_PAGE_IMAGE_FORMAT", "PNG")
 
 # yolox-page-elements-v1 contants
 YOLOX_PAGE_V1_NUM_CLASSES = 4
@@ -239,15 +238,11 @@ class YoloxModelInterfaceBase(ModelInterface):
                 # Convert to uint8 if needed.
                 if image.dtype != np.uint8:
                     image = (image * 255).astype(np.uint8)
-                # Convert the numpy array to a PIL Image.
-                image_pil = Image.fromarray(image)
-                original_size = image_pil.size
-
-                # Save the image to a buffer and encode to base64.
-                buffered = io.BytesIO()
-                image_pil.save(buffered, format="PNG")
-                image_b64 = base64.b64encode(buffered.getvalue()).decode("utf-8")
 
+                # Get original size directly from numpy array (width, height)
+                original_size = (image.shape[1], image.shape[0])
+                # Convert numpy array directly to base64 using OpenCV
+                image_b64 = numpy_to_base64(image, format=YOLOX_PAGE_IMAGE_FORMAT)
                 # Scale the image if necessary.
                 scaled_image_b64, new_size = scale_image_to_encoding_size(
                     image_b64, max_base64_size=self.nim_max_image_size

nv_ingest_api/util/image_processing/transforms.py

@@ -2,29 +2,52 @@
 # All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-import base64
-import io
 import logging
-from io import BytesIO
 from math import ceil
 from math import floor
 from typing import Optional
 from typing import Tuple
 
+import cv2
 import numpy as np
+from io import BytesIO
 from PIL import Image
-from PIL import UnidentifiedImageError
 
 from nv_ingest_api.util.converters import bytetools
 
+# Configure OpenCV to use a single thread for image processing
+cv2.setNumThreads(1)
 DEFAULT_MAX_WIDTH = 1024
 DEFAULT_MAX_HEIGHT = 1280
 
 logger = logging.getLogger(__name__)
 
 
+def _resize_image_opencv(
+    array: np.ndarray, target_size: Tuple[int, int], interpolation=cv2.INTER_LANCZOS4
+) -> np.ndarray:
+    """
+    Resizes a NumPy array representing an image using OpenCV.
+
+    Parameters
+    ----------
+    array : np.ndarray
+        The input image as a NumPy array.
+    target_size : Tuple[int, int]
+        The target size as (width, height).
+    interpolation : int, optional
+        OpenCV interpolation method. Defaults to cv2.INTER_LANCZOS4.
+
+    Returns
+    -------
+    np.ndarray
+        The resized image as a NumPy array.
+    """
+    return cv2.resize(array, target_size, interpolation=interpolation)
+
+
 def scale_image_to_encoding_size(
-    base64_image: str, max_base64_size: int = 180_000, initial_reduction: float = 0.9
+    base64_image: str, max_base64_size: int = 180_000, initial_reduction: float = 0.9, format: str = "PNG", **kwargs
 ) -> Tuple[str, Tuple[int, int]]:
     """
     Decodes a base64-encoded image, resizes it if needed, and re-encodes it as base64.
@@ -38,12 +61,19 @@ def scale_image_to_encoding_size(
         Maximum allowable size for the base64-encoded image, by default 180,000 characters.
     initial_reduction : float, optional
         Initial reduction step for resizing, by default 0.9.
+    format : str, optional
+        The image format to use for encoding. Supported formats are "PNG" and "JPEG".
+        Defaults to "PNG".
+    **kwargs
+        Additional keyword arguments passed to the format-specific encoding function.
+        For JPEG: quality (int, default=100) - JPEG quality (1-100).
+        For PNG: compression (int, default=3) - PNG compression level (0-9).
 
     Returns
     -------
     Tuple[str, Tuple[int, int]]
         A tuple containing:
-        - Base64-encoded PNG image string, resized if necessary.
+        - Base64-encoded image string in the specified format, resized if necessary.
         - The new size as a tuple (width, height).
 
     Raises
@@ -52,12 +82,11 @@ def scale_image_to_encoding_size(
         If the image cannot be resized below the specified max_base64_size.
     """
     try:
-        # Decode the base64 image and open it as a PIL image
-        image_data = base64.b64decode(base64_image)
-        img = Image.open(io.BytesIO(image_data)).convert("RGB")
+        # Decode the base64 image using OpenCV (returns RGB format)
+        img_array = base64_to_numpy(base64_image)
 
-        # Initial image size
-        original_size = img.size
+        # Initial image size (height, width, channels) -> (width, height)
+        original_size = (img_array.shape[1], img_array.shape[0])
 
         # Check initial size
         if len(base64_image) <= max_base64_size:
@@ -66,23 +95,24 @@ def scale_image_to_encoding_size(
         # Initial reduction step
         reduction_step = initial_reduction
         new_size = original_size
+        current_img = img_array.copy()
+        original_width, original_height = original_size
+
         while len(base64_image) > max_base64_size:
-            width, height = img.size
-            new_size = (int(width * reduction_step), int(height * reduction_step))
+            new_size = (int(original_width * reduction_step), int(original_height * reduction_step))
+            if new_size[0] < 1 or new_size[1] < 1:
+                raise ValueError("Image cannot be resized further without becoming too small.")
+
+            # Resize the image using OpenCV
+            current_img = _resize_image_opencv(img_array, new_size)
 
-            img_resized = img.resize(new_size, Image.LANCZOS)
-            buffered = io.BytesIO()
-            img_resized.save(buffered, format="PNG")
-            base64_image = base64.b64encode(buffered.getvalue()).decode("utf-8")
+            # Re-encode as base64 using the specified format
+            base64_image = numpy_to_base64(current_img, format=format, **kwargs)
 
             # Adjust the reduction step if necessary
             if len(base64_image) > max_base64_size:
                 reduction_step *= 0.95  # Reduce size further if needed
 
-            # Safety check
-            if new_size[0] < 1 or new_size[1] < 1:
-                raise Exception("Image cannot be resized further without becoming too small.")
-
         return base64_image, new_size
 
     except Exception as e:
@@ -90,36 +120,84 @@ def scale_image_to_encoding_size(
         raise
 
 
-def ensure_base64_is_png(base64_image: str) -> str:
+def _detect_base64_image_format(base64_string: str) -> Optional[str]:
     """
-    Ensures the given base64-encoded image is in PNG format. Converts to PNG if necessary.
+    Detects the format of a base64-encoded image using Pillow.
 
     Parameters
     ----------
-    base64_image : str
+    base64_string : str
         Base64-encoded image string.
 
     Returns
     -------
-    str
-        Base64-encoded PNG image string.
+    The detected format ("PNG", "JPEG", "UNKNOWN")
     """
     try:
-        # Decode the base64 string and load the image
-        image_data = base64.b64decode(base64_image)
-        image = Image.open(io.BytesIO(image_data))
+        image_bytes = bytetools.bytesfrombase64(base64_string)
+    except Exception as e:
+        logger.error(f"Invalid base64 string: {e}")
+        raise ValueError(f"Invalid base64 string: {e}") from e
+
+    try:
+        with Image.open(BytesIO(image_bytes)) as img:
+            return img.format.upper()
+    except ImportError:
+        raise ImportError("Pillow library not available")
+    except Exception as e:
+        logger.error(f"Error detecting image format: {e}")
+        return "UNKNOWN"
 
-        # Check if the image is already in PNG format
-        if image.format != "PNG":
-            # Convert the image to PNG
-            buffered = io.BytesIO()
-            image.convert("RGB").save(buffered, format="PNG")
-            base64_image = base64.b64encode(buffered.getvalue()).decode("utf-8")
 
+def ensure_base64_format(base64_image: str, target_format: str = "PNG", **kwargs) -> str:
+    """
+    Ensures the given base64-encoded image is in the specified format. Converts if necessary.
+    Skips conversion if the image is already in the target format.
+
+    Parameters
+    ----------
+    base64_image : str
+        Base64-encoded image string.
+    target_format : str, optional
+        The target image format. Supported formats are "PNG" and "JPEG". Defaults to "PNG".
+    **kwargs
+        Additional keyword arguments passed to the format-specific encoding function.
+        For JPEG: quality (int, default=100) - JPEG quality (1-100).
+        For PNG: compression (int, default=3) - PNG compression level (0-9).
+
+    Returns
+    -------
+    str
+        Base64-encoded image string in the specified format.
+
+    Raises
+    ------
+    ValueError
+        If there is an error during format conversion.
+    """
+    target_format = target_format.upper()
+    if target_format == "JPG":
+        target_format = "JPEG"
+
+    current_format = _detect_base64_image_format(base64_image)
+    if current_format == "UNKNOWN":
+        raise ValueError(
+            f"Unable to decode image from base64 string: {base64_image}, because current format could not be detected."
+        )
+    if current_format == target_format:
+        logger.debug(f"Image already in {target_format} format, skipping conversion")
         return base64_image
+
+    try:
+        # Decode the base64 image using OpenCV (returns RGB format)
+        img_array = base64_to_numpy(base64_image)
+        # Re-encode in the target format
+        return numpy_to_base64(img_array, format=target_format, **kwargs)
+    except ImportError as e:
+        raise e
     except Exception as e:
-        logger.error(f"Error ensuring PNG format: {e}")
-        return None
+        logger.error(f"Error converting image to {target_format} format: {e}")
+        raise ValueError(f"Failed to convert image to {target_format} format: {e}") from e
 
 
 def pad_image(
@@ -302,66 +380,193 @@ def normalize_image(
     return output_array
 
 
-def numpy_to_base64(array: np.ndarray) -> str:
+def _preprocess_numpy_array(array: np.ndarray) -> np.ndarray:
+    """
+    Preprocesses a NumPy array for image encoding by ensuring proper format and data type.
+    Also handles color space conversion for OpenCV encoding.
+
+    Parameters
+    ----------
+    array : np.ndarray
+        The input image as a NumPy array.
+
+    Returns
+    -------
+    np.ndarray
+        The preprocessed array in uint8 format, ready for OpenCV encoding (BGR color order for color images).
+
+    Raises
+    ------
+    ValueError
+        If the input array cannot be converted into a valid image format.
+    """
+    # Check if the array is valid and can be converted to an image
+    try:
+        # If the array represents a grayscale image, drop the redundant axis in
+        # (h, w, 1). cv2 expects (h, w) for grayscale.
+        if array.ndim == 3 and array.shape[2] == 1:
+            array = np.squeeze(array, axis=2)
+
+        # Ensure uint8 data type
+        processed_array = array.astype(np.uint8)
+
+        # OpenCV uses BGR color order, so convert RGB to BGR if needed
+        if processed_array.ndim == 3 and processed_array.shape[2] == 3:
+            # Assume input is RGB and convert to BGR for OpenCV
+            processed_array = cv2.cvtColor(processed_array, cv2.COLOR_RGB2BGR)
+
+        return processed_array
+    except Exception as e:
+        raise ValueError(f"Failed to preprocess NumPy array for image encoding: {e}")
+
+
+def _encode_opencv_jpeg(array: np.ndarray, *, quality: int = 100) -> bytes:
+    """NumPy array -> JPEG bytes using OpenCV."""
+    ok, buf = cv2.imencode(".jpg", array, [int(cv2.IMWRITE_JPEG_QUALITY), quality])
+    if not ok:
+        raise RuntimeError("cv2.imencode failed")
+    return buf.tobytes()
+
+
+def _encode_opencv_png(array: np.ndarray, *, compression: int = 6) -> bytes:
+    """NumPy array -> PNG bytes using OpenCV"""
+    encode_params = [
+        cv2.IMWRITE_PNG_COMPRESSION,
+        compression,
+        cv2.IMWRITE_PNG_STRATEGY,
+        cv2.IMWRITE_PNG_STRATEGY_DEFAULT,
+    ]
+    ok, buf = cv2.imencode(".png", array, encode_params)
+    if not ok:
+        raise RuntimeError("cv2.imencode(.png) failed")
+    return buf.tobytes()
+
+
+def numpy_to_base64_png(array: np.ndarray) -> str:
+    """
+    Converts a preprocessed NumPy array representing an image to a base64-encoded PNG string using OpenCV.
+
+    Parameters
+    ----------
+    array : np.ndarray
+        The preprocessed input image as a NumPy array. Must have a shape compatible with image data.
+
+    Returns
+    -------
+    str
+        The base64-encoded PNG string representation of the input NumPy array.
+
+    Raises
+    ------
+    RuntimeError
+        If there is an issue during the image conversion or base64 encoding process.
+    """
+    try:
+        # Encode to PNG bytes using OpenCV
+        png_bytes = _encode_opencv_png(array)
+
+        # Convert to base64
+        base64_img = bytetools.base64frombytes(png_bytes)
+    except Exception as e:
+        raise RuntimeError(f"Failed to encode image to base64 PNG: {e}")
+
+    return base64_img
+
+
+def numpy_to_base64_jpeg(array: np.ndarray, quality: int = 100) -> str:
+    """
+    Converts a preprocessed NumPy array representing an image to a base64-encoded JPEG string using OpenCV.
+
+    Parameters
+    ----------
+    array : np.ndarray
+        The preprocessed input image as a NumPy array. Must have a shape compatible with image data.
+    quality : int, optional
+        JPEG quality (1-100), by default 100. Higher values mean better quality but larger file size.
+
+    Returns
+    -------
+    str
+        The base64-encoded JPEG string representation of the input NumPy array.
+
+    Raises
+    ------
+    RuntimeError
+        If there is an issue during the image conversion or base64 encoding process.
+    """
+    try:
+        # Encode to JPEG bytes using OpenCV
+        jpeg_bytes = _encode_opencv_jpeg(array, quality=quality)
+
+        # Convert to base64
+        base64_img = bytetools.base64frombytes(jpeg_bytes)
+    except Exception as e:
+        raise RuntimeError(f"Failed to encode image to base64 JPEG: {e}")
+
+    return base64_img
+
+
+def numpy_to_base64(array: np.ndarray, format: str = "PNG", **kwargs) -> str:
     """
     Converts a NumPy array representing an image to a base64-encoded string.
 
-    The function takes a NumPy array, converts it to a PIL image, and then encodes
-    the image as a PNG in a base64 string format. The input array is expected to be in
-    a format that can be converted to a valid image, such as having a shape of (H, W, C)
-    where C is the number of channels (e.g., 3 for RGB).
+    The function takes a NumPy array, preprocesses it, and then encodes
+    the image in the specified format as a base64 string. The input array is expected
+    to be in a format that can be converted to a valid image, such as having a shape
+    of (H, W, C) where C is the number of channels (e.g., 3 for RGB).
 
     Parameters
     ----------
     array : np.ndarray
         The input image as a NumPy array. Must have a shape compatible with image data.
+    format : str, optional
+        The image format to use for encoding. Supported formats are "PNG" and "JPEG".
+        Defaults to "PNG".
+    **kwargs
+        Additional keyword arguments passed to the format-specific encoding function.
+        For JPEG: quality (int, default=100) - JPEG quality (1-100).
 
     Returns
     -------
     str
-        The base64-encoded string representation of the input NumPy array as a PNG image.
+        The base64-encoded string representation of the input NumPy array in the specified format.
 
     Raises
     ------
     ValueError
-        If the input array cannot be converted into a valid image format.
+        If the input array cannot be converted into a valid image format, or if an
+        unsupported format is specified.
     RuntimeError
         If there is an issue during the image conversion or base64 encoding process.
 
     Examples
     --------
     >>> array = np.random.randint(0, 255, (100, 100, 3), dtype=np.uint8)
-    >>> encoded_str = numpy_to_base64(array)
+    >>> encoded_str = numpy_to_base64(array, format="PNG")
     >>> isinstance(encoded_str, str)
     True
+    >>> encoded_str_jpeg = numpy_to_base64(array, format="JPEG", quality=90)
+    >>> isinstance(encoded_str_jpeg, str)
+    True
     """
-    # If the array represents a grayscale image, drop the redundant axis in
-    # (h, w, 1). PIL.Image.fromarray() expects an array of form (h, w) if it's
-    # a grayscale image.
-    if array.ndim == 3 and array.shape[2] == 1:
-        array = np.squeeze(array, axis=2)
+    # Centralized preprocessing of the numpy array
+    processed_array = _preprocess_numpy_array(array)
 
-    # Check if the array is valid and can be converted to an image
-    try:
-        # Convert the NumPy array to a PIL image
-        pil_image = Image.fromarray(array.astype(np.uint8))
-    except Exception as e:
-        raise ValueError(f"Failed to convert NumPy array to image: {e}")
+    format = format.upper()
 
-    try:
-        # Convert the PIL image to a base64-encoded string
-        with BytesIO() as buffer:
-            pil_image.save(buffer, format="PNG")
-            base64_img = bytetools.base64frombytes(buffer.getvalue())
-    except Exception as e:
-        raise RuntimeError(f"Failed to encode image to base64: {e}")
-
-    return base64_img
+    if format == "PNG":
+        return numpy_to_base64_png(processed_array)
+    elif format == "JPEG" or format == "JPG":
+        quality = kwargs.get("quality", 100)
+        return numpy_to_base64_jpeg(processed_array, quality=quality)
+    else:
+        raise ValueError(f"Unsupported format: {format}. Supported formats are 'PNG' and 'JPEG'.")
 
 
 def base64_to_numpy(base64_string: str) -> np.ndarray:
     """
-    Convert a base64-encoded image string to a NumPy array.
+    Convert a base64-encoded image string to a NumPy array using OpenCV.
+    Returns images in RGB format for consistency.
 
     Parameters
     ----------
@@ -371,37 +576,82 @@ def base64_to_numpy(base64_string: str) -> np.ndarray:
     Returns
     -------
     numpy.ndarray
-        NumPy array representation of the decoded image.
+        NumPy array representation of the decoded image in RGB format (for color images).
+        Grayscale images are returned as-is.
 
     Raises
     ------
     ValueError
         If the base64 string is invalid or cannot be decoded into an image.
-    ImportError
-        If required libraries are not installed.
 
     Examples
     --------
     >>> base64_str = '/9j/4AAQSkZJRgABAQAAAQABAAD/2wBD...'
     >>> img_array = base64_to_numpy(base64_str)
+    >>> # img_array is now in RGB format (for color images)
     """
     try:
-        # Decode the base64 string
-        image_data = base64.b64decode(base64_string)
-    except (base64.binascii.Error, ValueError) as e:
+        # Decode the base64 string to bytes using bytetools
+        image_bytes = bytetools.bytesfrombase64(base64_string)
+    except Exception as e:
         raise ValueError("Invalid base64 string") from e
 
+    # Create numpy buffer from bytes and decode using OpenCV
+    buf = np.frombuffer(image_bytes, dtype=np.uint8)
     try:
-        # Convert the bytes into a BytesIO object
-        image_bytes = BytesIO(image_data)
-
-        # Open the image using PIL
-        image = Image.open(image_bytes)
-        image.load()
-    except UnidentifiedImageError as e:
+        img = cv2.imdecode(buf, cv2.IMREAD_UNCHANGED)
+        if img is None:
+            raise ValueError("OpenCV failed to decode image")
+
+        # Convert BGR to RGB for consistent processing (OpenCV loads as BGR)
+        # Only convert if it's a 3-channel color image
+        if img.ndim == 3 and img.shape[2] == 3:
+            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+    except ImportError:
+        raise
+    except Exception as e:
         raise ValueError("Unable to decode image from base64 string") from e
 
-    # Convert the image to a NumPy array
-    image_array = np.array(image)
+    # Convert to numpy array
+    img = np.array(img)
+    # Assert that 3-channel images are in RGB format after conversion
+    assert img.ndim <= 3, f"Image has unexpected number of dimensions: {img.ndim}"
+    assert img.ndim != 3 or img.shape[2] == 3, f"3-channel image should have 3 channels, got: {img.shape[2]}"
+
+    return img
+
+
+def scale_numpy_image(
+    img_arr: np.ndarray, scale_tuple: Optional[Tuple[int, int]] = None, interpolation=Image.LANCZOS
+) -> np.ndarray:
+    """
+    Scales a NumPy image array using OpenCV with aspect ratio preservation.
 
-    return image_array
+    This function provides OpenCV-based image scaling that mimics PIL's thumbnail behavior
+    by maintaining aspect ratio and scaling to fit within the specified dimensions.
+
+    Parameters
+    ----------
+    img_arr : np.ndarray
+        The input image as a NumPy array.
+    scale_tuple : Optional[Tuple[int, int]], optional
+        A tuple (width, height) to resize the image to. If provided, the image
+        will be resized to fit within these dimensions while maintaining aspect ratio
+        (similar to PIL's thumbnail method). Defaults to None.
+    interpolation : int, optional
+        OpenCV interpolation method. Defaults to cv2.INTER_LANCZOS4.
+
+    Returns
+    -------
+    np.ndarray
+        A NumPy array representing the scaled image data.
+    """
+    # Apply scaling using OpenCV if specified
+    # Using PIL for scaling as CV2 seems to lead to different results
+    # TODO: Remove when we move to YOLOX Ensemble Models
+    if scale_tuple:
+        image = Image.fromarray(img_arr)
+        image.thumbnail(scale_tuple, interpolation)
+        img_arr = np.array(image)
+    # Ensure we return a copy
+    return img_arr.copy()

nv_ingest_api/util/pdf/pdfium.py

@@ -7,7 +7,6 @@ from typing import List, Any
 from typing import Optional
 from typing import Tuple
 
-import PIL
 import numpy as np
 import pypdfium2 as pdfium
 import pypdfium2.raw as pdfium_c
@@ -20,8 +19,9 @@ from nv_ingest_api.util.image_processing.clustering import (
     combine_groups_into_bboxes,
     remove_superset_bboxes,
 )
-from nv_ingest_api.util.image_processing.transforms import pad_image, numpy_to_base64, crop_image
+from nv_ingest_api.util.image_processing.transforms import pad_image, numpy_to_base64, crop_image, scale_numpy_image
 from nv_ingest_api.util.metadata.aggregators import Base64Image
+from nv_ingest_api.internal.primitives.nim.model_interface.yolox import YOLOX_PAGE_IMAGE_FORMAT
 
 logger = logging.getLogger(__name__)
 
@@ -176,18 +176,10 @@ def pdfium_pages_to_numpy(
     for idx, page in enumerate(pages):
         # Render the page as a bitmap with the specified scale and rotation
         page_bitmap = page.render(scale=scale, rotation=rotation)
-
-        # Convert the bitmap to a PIL image
-        pil_image = page_bitmap.to_pil()
-
+        img_arr = convert_bitmap_to_corrected_numpy(page_bitmap)
         # Apply scaling using the thumbnail approach if specified
         if scale_tuple:
-            pil_image.thumbnail(scale_tuple, PIL.Image.LANCZOS)
-
-        # Convert the PIL image to a NumPy array and force a full copy,
-        # ensuring the returned array is entirely independent of the original buffer.
-        img_arr = np.array(pil_image).copy()
-
+            img_arr = scale_numpy_image(img_arr, scale_tuple)
         # Apply padding if specified
         if padding_tuple:
             img_arr, (pad_width, pad_height) = pad_image(
@@ -250,7 +242,7 @@ def extract_simple_images_from_pdfium_page(page, max_depth):
         try:
             # Attempt to retrieve the image bitmap
             image_numpy: np.ndarray = pdfium_try_get_bitmap_as_numpy(obj)  # noqa
-            image_base64: str = numpy_to_base64(image_numpy)
+            image_base64: str = numpy_to_base64(image_numpy, format=YOLOX_PAGE_IMAGE_FORMAT)
             image_bbox = obj.get_pos()
             image_size = obj.get_size()
             if image_size[0] < 10 and image_size[1] < 10:

{nv_ingest_api-2025.7.13.dev20250713.dist-info → nv_ingest_api-2025.7.15.dev20250715.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nv-ingest-api
-Version: 2025.7.13.dev20250713
+Version: 2025.7.15.dev20250715
 Summary: Python module with core document ingestion functions.
 Author-email: Jeremy Dyer <jdyer@nvidia.com>
 License:                                  Apache License
@@ -217,6 +217,7 @@ Requires-Dist: backoff==2.2.1
 Requires-Dist: pandas>=2.0
 Requires-Dist: pydantic>2.0.0
 Requires-Dist: pydantic-settings>2.0.0
+Requires-Dist: tritonclient
 Dynamic: license-file
 
 # nv-ingest-api

{nv_ingest_api-2025.7.13.dev20250713.dist-info → nv_ingest_api-2025.7.15.dev20250715.dist-info}/RECORD

@@ -31,8 +31,8 @@ nv_ingest_api/internal/extract/pdf/pdf_extractor.py,sha256=CxtWaD6mql9MEqSdk2CfS
 nv_ingest_api/internal/extract/pdf/engines/__init__.py,sha256=u4GnAZmDKRl0RwYGIRiozIRw70Kybw3A72-lcKFeoTI,582
 nv_ingest_api/internal/extract/pdf/engines/adobe.py,sha256=VT0dEqkU-y2uGkaCqxtKYov_Q8R1028UQVBchgMLca4,17466
 nv_ingest_api/internal/extract/pdf/engines/llama.py,sha256=PpKTqS8jGHBV6mKLGZWwjpfT8ga6Fy8ffrvL-gPAf2c,8182
-nv_ingest_api/internal/extract/pdf/engines/nemoretriever.py,sha256=Uqj1NH7yWga9P6_vCzgny1WKALfF--UdAaGHUF8K_aQ,22926
-nv_ingest_api/internal/extract/pdf/engines/pdfium.py,sha256=fDbrZwJ-lgeHYOq107WXehzdSvyF8zEDza_9UkDm5aE,22360
+nv_ingest_api/internal/extract/pdf/engines/nemoretriever.py,sha256=XNYz4S2tMFBv0KFzXNERrVs-1raxJ_iIIXpBGlJFcD0,22987
+nv_ingest_api/internal/extract/pdf/engines/pdfium.py,sha256=vtdBue1EEQJsHcBuX3NdPutbLfyKPIzily6JOK6yV0w,22421
 nv_ingest_api/internal/extract/pdf/engines/tika.py,sha256=6GyR2l6EsgNZl9jnYDXLeKNK9Fj2Mw9y2UWDq-eSkOc,3169
 nv_ingest_api/internal/extract/pdf/engines/unstructured_io.py,sha256=jrv2B4VZAH4PevAQrFz965qz8UyXq3rViiOTbGLejec,14908
 nv_ingest_api/internal/extract/pdf/engines/pdf_helpers/__init__.py,sha256=Jk3wrQ2CZs167juvEZ-uV6qXWQjR08hhIu8otk2MWj4,4931
@@ -55,12 +55,12 @@ nv_ingest_api/internal/primitives/nim/model_interface/cached.py,sha256=b1HX-PY1E
 nv_ingest_api/internal/primitives/nim/model_interface/decorators.py,sha256=qwubkHs4WjnexM6rI0wkjWCsrVNEbA4Wjk2oKL9OYCU,1499
 nv_ingest_api/internal/primitives/nim/model_interface/deplot.py,sha256=TvKdk6PTuI1WNhRmNNrvygaI_DIutkJkDL-XdtLZQac,10787
 nv_ingest_api/internal/primitives/nim/model_interface/helpers.py,sha256=x35a9AyTYxpESQflLo_YnhVOKblQKVen6vGGFaXmNiE,9927
-nv_ingest_api/internal/primitives/nim/model_interface/nemoretriever_parse.py,sha256=MFWPqMTXs_MZG3ripRR21o7f_mVeoE46Q10yvJ8KNr0,7023
+nv_ingest_api/internal/primitives/nim/model_interface/nemoretriever_parse.py,sha256=WysjDZeegclO3mZgVcGOwzWbr8wSI4pWRiYD4iC2EXo,7098
 nv_ingest_api/internal/primitives/nim/model_interface/paddle.py,sha256=rSUPwl5XOrqneoS6aKhatVjrNBg_LhP3nwUWS_aTwz0,17950
 nv_ingest_api/internal/primitives/nim/model_interface/parakeet.py,sha256=5PqD2JuHY2rwd-6SSB4axr2Dd79vm95sAEkcmI3U7ME,12977
 nv_ingest_api/internal/primitives/nim/model_interface/text_embedding.py,sha256=lFhppNqrq5X_fzbCWKphvZQMzaJd3gHrkWsyJORzFrU,5010
 nv_ingest_api/internal/primitives/nim/model_interface/vlm.py,sha256=qJ382PU1ZrIM-SR3cqIhtY_W2rmHec2HIa2aUB2SvaU,6031
-nv_ingest_api/internal/primitives/nim/model_interface/yolox.py,sha256=uYXqdvqgkyS4Yfr9ZoikRDX4e94OV3ch3Xhv3JVg-3s,49581
+nv_ingest_api/internal/primitives/nim/model_interface/yolox.py,sha256=nsfDQgeupBe9Tdf3S5sfNpYcObEwVlzCZdfg1ObAW88,49584
 nv_ingest_api/internal/primitives/tracing/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nv_ingest_api/internal/primitives/tracing/latency.py,sha256=5kVTeYRbRdTlT_aI4MeS20N_S7mqCcLqZR6YHtxhXkY,2215
 nv_ingest_api/internal/primitives/tracing/logging.py,sha256=SSzIgS7afLH-e1C7VagYDmkkA6rTXmQ-bmtLjoEguhg,3851
@@ -123,7 +123,7 @@ nv_ingest_api/util/image_processing/__init__.py,sha256=Jiy8C1ZuSrNb_eBM1ZTV9IKFI
 nv_ingest_api/util/image_processing/clustering.py,sha256=sUGlZI4cx1q8h4Pns1N9JVpdfSM2BOH8zRmn9QFCtzI,9236
 nv_ingest_api/util/image_processing/processing.py,sha256=LSoDDEmahr7a-qSS12McVcowRe3dOrAZwa1h-PD_JPQ,6554
 nv_ingest_api/util/image_processing/table_and_chart.py,sha256=bxOu9PZYkG_WFCDGw_JLaO60S2pDSN8EOWK3xkIwr2A,14376
-nv_ingest_api/util/image_processing/transforms.py,sha256=Kz9hrizV314Hy7cRCYK9ZmhmBbVUOZ_z0HEpzZYcslQ,14081
+nv_ingest_api/util/image_processing/transforms.py,sha256=CJVGQgUvHk_mzihR8ZZrvwJUBgUYcgFAKzXyRTmKdCE,23371
 nv_ingest_api/util/imports/__init__.py,sha256=wQSlVx3T14ZgQAt-EPzEczQusXVW0W8yynnUaFFGE3s,143
 nv_ingest_api/util/imports/callable_signatures.py,sha256=e2bJB1pmkN4Ee-Bf-VggOSBaQ4RXofWF5eKkWXgIj2U,1855
 nv_ingest_api/util/imports/dynamic_resolvers.py,sha256=7GByV_-8z2X0tnVoabCxVioxOP3sYMros3ZllVAW-wY,4343
@@ -140,7 +140,7 @@ nv_ingest_api/util/multi_processing/__init__.py,sha256=4fojP8Rp_5Hu1YAkqGylqTyEZ
 nv_ingest_api/util/multi_processing/mp_pool_singleton.py,sha256=dTfP82DgGPaXEJH3jywTO8rNlLZUniD4FFzwv84_giE,7372
 nv_ingest_api/util/nim/__init__.py,sha256=UqbiXFCqjWcjNvoduXd_0gOUOGBT8JvppiYHOmMyneA,1775
 nv_ingest_api/util/pdf/__init__.py,sha256=uLsBITo_XfgbwpzqXUm1IYX6XlZrTfx6T1cIhdILwG8,140
-nv_ingest_api/util/pdf/pdfium.py,sha256=Ch9Gh5jRLcBr3stjCckqWwTUL-T0sI50PlQnZHo_9NA,15761
+nv_ingest_api/util/pdf/pdfium.py,sha256=qTiTlSaiCk_rxm_eoQBoAFKq_5OQrioHVSbPbGDxVkE,15668
 nv_ingest_api/util/schema/__init__.py,sha256=wQSlVx3T14ZgQAt-EPzEczQusXVW0W8yynnUaFFGE3s,143
 nv_ingest_api/util/schema/schema_validator.py,sha256=H0yZ_i_HZaiBRUCGmTBfRB9-hURhVqyd10aS_ynM1_0,321
 nv_ingest_api/util/service_clients/__init__.py,sha256=wQSlVx3T14ZgQAt-EPzEczQusXVW0W8yynnUaFFGE3s,143
@@ -153,8 +153,8 @@ nv_ingest_api/util/service_clients/rest/rest_client.py,sha256=dZ-jrk7IK7oNtHoXFS
 nv_ingest_api/util/string_processing/__init__.py,sha256=mkwHthyS-IILcLcL1tJYeF6mpqX3pxEw5aUzDGjTSeU,1411
 nv_ingest_api/util/system/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nv_ingest_api/util/system/hardware_info.py,sha256=ORZeKpH9kSGU_vuPhyBwkIiMyCViKUX2CP__MCjrfbU,19463
-nv_ingest_api-2025.7.13.dev20250713.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-nv_ingest_api-2025.7.13.dev20250713.dist-info/METADATA,sha256=MMMCpLepD3X5G3COkoHsDpO9CkaoL_DdFuE_t4OsPAk,13919
-nv_ingest_api-2025.7.13.dev20250713.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-nv_ingest_api-2025.7.13.dev20250713.dist-info/top_level.txt,sha256=abjYMlTJGoG5tOdfIB-IWvLyKclw6HLaRSc8MxX4X6I,14
-nv_ingest_api-2025.7.13.dev20250713.dist-info/RECORD,,
+nv_ingest_api-2025.7.15.dev20250715.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+nv_ingest_api-2025.7.15.dev20250715.dist-info/METADATA,sha256=OWZyeCR9DZ23SdT0RcMdodCkxR508CZZaVczdM3qXPE,13947
+nv_ingest_api-2025.7.15.dev20250715.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+nv_ingest_api-2025.7.15.dev20250715.dist-info/top_level.txt,sha256=abjYMlTJGoG5tOdfIB-IWvLyKclw6HLaRSc8MxX4X6I,14
+nv_ingest_api-2025.7.15.dev20250715.dist-info/RECORD,,