非缩放方式降低图像分辨率的技术咨询

Hey there, let's clear up a key misunderstanding first—DPI has nothing to do with an image's actual pixel resolution! That's why tweaking it didn't give you the result you wanted, and why you got a KeyError when accessing im.info['dpi'] (some images don't store DPI metadata at all, which is totally normal). DPI is just a print-time parameter that tells printers how many pixels to fit per inch; it doesn't change the number of pixels in your image.

真正的降分辨率方法：下采样（Downsampling） #

To reduce the image's pixel count (which is what you actually need for target detection), you'll want to use downsampling. Here are two practical approaches, including the blur+decimate method mentioned in the comments:

方法1：直接使用Pillow的resize（带抗锯齿插值）

This is the simplest and most common way—it resizes the image to your target dimensions using a high-quality interpolation method to avoid jagged edges:

from PIL import Image

# 打开原始图像
im = Image.open("car.png")

# 定义目标分辨率：比如降到原尺寸的1/2（宽高各除以2）
target_width = im.width // 2
target_height = im.height // 2

# 使用LANCZOS插值（抗锯齿效果最优）进行降采样
low_res_im = im.resize((target_width, target_height), Image.Resampling.LANCZOS)

# 保存结果
low_res_im.save("low_res_car.png")

方法2：先高斯模糊再抽取（Decimate）

This method addresses aliasing (jagged artifacts) by first blurring the image to smooth high-frequency details, then "decimating" (sampling every N-th pixel) to reduce resolution. It's great for preserving edge clarity in target detection:

from PIL import Image, ImageFilter
import numpy as np

# 打开图像并先高斯模糊
im = Image.open("car.png")
# 模糊半径根据下采样比例调整：下采样2倍用半径1，下采样4倍用半径2，以此类推
blurred_im = im.filter(ImageFilter.GaussianBlur(radius=1))

# 方式A：用resize配合NEAREST插值实现抽取
decimated_im = blurred_im.resize((blurred_im.width//2, blurred_im.height//2), Image.Resampling.NEAREST)

# 方式B：用numpy直接抽取像素（更高效，适合批量处理）
im_array = np.array(blurred_im)
# 每隔2个像素取一个（行和列都取步长为2的索引）
decimated_array = im_array[::2, ::2, :]
decimated_im = Image.fromarray(decimated_array)

# 保存结果
decimated_im.save("blur_decimate_car.png")

为什么先模糊再抽取？ #

直接 decimating (sampling every N-th pixel) can cause aliasing—jagged lines or moiré patterns from high-frequency details. Gaussian blur smooths those details first, ensuring the downsampled image remains clean and suitable for target detection models.

Glad to hear you've already solved the problem with two approaches! These methods should align with what you implemented.

内容的提问来源于stack exchange，提问作者user13713033