Implementation:Kserve Kserve Paddle Image Preprocess

Knowledge Sources	Kserve_Kserve
Domains	Image Processing, PaddlePaddle
Last Updated	2026-02-13 00:00 GMT

Overview

Concrete tool for preprocessing images for PaddlePaddle ResNet50 inference with resizing, center cropping, and ImageNet normalization provided by the KServe sample code.

Description

This module provides three image preprocessing functions designed for PaddlePaddle model inference:

• resize_short() -- Resizes an image so that its shortest side matches the target size while preserving the aspect ratio. Uses OpenCV's cv2.resize().
• crop_image() -- Crops an image to a square of target_size. If center is True, performs a center crop; otherwise, performs a random crop.
• preprocess() -- Combines the full preprocessing pipeline: resizes the image to 224 on the short side, center-crops to 224x224, converts BGR to RGB, transposes from HWC to CHW format, normalizes to [0, 1], and applies ImageNet mean ([0.485, 0.456, 0.406]) and standard deviation ([0.229, 0.224, 0.225]) normalization. Returns the result with a batch dimension added.

Usage

Use these functions to preprocess images before sending them to a PaddlePaddle ResNet50 model deployed on KServe, ensuring consistency with the training-time transformations.

Code Reference

Source Location

• Repository: Kserve_Kserve
• File: docs/samples/v1beta1/paddle/img_preprocess.py
• Lines: 1-41

Signature

def resize_short(img, target_size):
    """resize_short"""
    ...

def crop_image(img, target_size, center):
    """crop_image"""
    ...

def preprocess(img):
    ...

Import

from img_preprocess import resize_short, crop_image, preprocess

I/O Contract

Inputs

resize_short()

Name	Type	Required	Description
img	np.ndarray	Yes	Input image as a NumPy array (HWC format)
target_size	int	Yes	Target size for the shortest side of the image

crop_image()

Name	Type	Required	Description
img	np.ndarray	Yes	Input image as a NumPy array (HWC format)
target_size	int	Yes	Size of the square crop
center	bool	Yes	If True, performs center crop; if False, performs random crop

preprocess()

Name	Type	Required	Description
img	np.ndarray	Yes	Input image as a NumPy array in BGR format (HWC layout)

Outputs

resize_short()

Name	Type	Description
resized	np.ndarray	Resized image with shortest side equal to target_size

crop_image()

Name	Type	Description
img	np.ndarray	Cropped image of shape (target_size, target_size, channels)

preprocess()

Name	Type	Description
img	np.ndarray	Preprocessed image of shape (1, 3, 224, 224) with ImageNet normalization applied

Usage Examples

Basic Usage

import cv2
from img_preprocess import preprocess

# Load an image using OpenCV (returns BGR format)
img = cv2.imread("sample_image.jpg")

# Preprocess for PaddlePaddle ResNet50
processed = preprocess(img)
print(processed.shape)  # (1, 3, 224, 224)

# Send to KServe prediction endpoint
import requests
payload = {"instances": processed.tolist()}
response = requests.post("http://localhost:8080/v1/models/resnet50:predict", json=payload)

Related Pages

• Environment:Kserve_Kserve_Python_Runtime