How Jev Makes Decisions About Images, With 4 Python Examples

%pip install -U "typesafe-sdk[pil]"

from getpass import getpass
from PIL import Image as PILImage, ImageOps
from typesafe_sdk import (
TypeSafeClient,
Image as JevImage,
Choice,
Noul,
Score,
Box,
)
client = TypeSafeClient(
api_key=getpass("TypeSafe API key: "),
model="jev-1.14.0",
headers={"TypeSafe-Danger-Zone-Images": "true"},
)
def load_photo(path):
"""Rotate upright and fit within a square without stretching."""
with PILImage.open(path) as original:
upright = ImageOps.exif_transpose(original).convert("RGB")
return ImageOps.pad(
upright,
(1024, 1024),
color="white",
)
def image_asset(path):
return JevImage.from_pil(load_photo(path), format="jpeg")

response = client.system_one(
state={
"sample": image_asset(r"d:\images\jev-test\to_classify.png"),
},
questions={
"pattern": Choice(
instructions=(
"哪张参考图案最符合 `sample` 中的织物？比较图案结构，忽略 "
"颜色、光照和拍摄角度的差异。"
),
criteria={
"striped": {
"description": "重复的平行条纹。",
"reference": image_asset(r"d:\images\jev-test\stripes.png"),
},
"checked": {
"description": "重复的格子图案。",
"reference": image_asset(r"d:\images\jev-test\check.png"),
},
"floral": {
"description": "重复的花卉或植物图案。",
"reference": image_asset(r"d:\images\jev-test\floral.png"),
},
"other": (
"纯色织物、其他图案、无可见织物，"
"或细节不足无法识别图案。"
),
},
),
},
)
answer = response.choices["pattern"]
print("Pattern:", answer.choice)
print("Confidence:", answer.confidence)
print("Option probabilities:", answer.probabilities)
if answer.choice == "other" or answer.confidence < 0.80:
print("Action: send to the catalogue review queue")
else:
print(f"Action: suggest the catalogue tag '{answer.choice}'")

Pattern: striped
Confidence: 1.0
Option probabilities: {'striped': 1.0, 'checked': 0.0, 'floral': 0.0, 'other': 0.0}
Action: suggest the catalogue tag 'striped'

Pattern: other
Confidence: 1.0
Option probabilities: {'striped': 0.0, 'checked': 0.0, 'floral': 0.0, 'other': 1.0}
Action: send to the catalogue review queue

images = {
"photo_1": r"d:\images\jev-test\product_photo_1.png",
"photo_2": r"d:\images\jev-test\product_photo_2.png",
}
criteria = [
"严重模糊：产品边缘难以区分。",
"明显模糊：边缘可见但细节丢失。",
"基本清晰：边缘清晰但存在轻微模糊的细部。",
"清晰：边缘和可见细部都清晰可辨。",
]
questions = {}
for name in images:
questions[f"{name}_visible"] = Noul(
instructions=(
f"在 `{name}` 中是否能明确识别出单一主体产品？ "
"仅评估当前图片。"
),
)
questions[f"{name}_sharpness"] = Score(
instructions=(
f"评估 `{name}` 中主体产品的对焦清晰度。 "
"仅评估当前图片。忽略有意的背景模糊。 "
"判断可见的产品边缘和表面细节。"
),
criteria=criteria,
)
response = client.system_one(
state={
name: image_asset(path)
for name, path in images.items()
},
questions=questions,
)
for name, path in images.items():
visible = response.nouls[f"{name}_visible"].noul
sharpness = response.scores[f"{name}_sharpness"]
print(f"\nImage: {path}")
print(f"产品可见概率: {visible:.3f}")
print(f"清晰度: {sharpness.score:.2f} / 3")
print(f"置信度: {sharpness.confidence:.3f}")
print("等级概率:", sharpness.probabilities)
if visible < 0.85:
print("操作: 要求提供一张主体明确的照片")
elif sharpness.confidence < 0.80:
print("操作: 人工审核照片")
elif sharpness.score < 2.0:
print("操作: 建议重新拍摄")
else:
print("操作: 通过对焦检查")

Image: d:\images\jev-test\product_photo_1.png
产品可见概率: 0.990
清晰度: 2.88 / 3
置信度: 0.870
等级概率: {0: 0.0, 1: 0.01, 2: 0.11, 3: 0.88}
操作: 通过对焦检查
Image: d:\images\jev-test\product_photo_2.png
产品可见概率: 0.970
清晰度: 1.04 / 3
置信度: 0.870
等级概率: {0: 0.04, 1: 0.88, 2: 0.07, 3: 0.01}
操作: 建议重新拍摄

from pathlib import Path
from math import floor, ceil
# 保留准备好的图像：返回的坐标引用此版本。
photo = load_photo(r"d:\images\jev-test\parcel_photo.png")
response = client.system_one(
state={
"parcel_photo": JevImage.from_pil(photo, format="png"),
},
questions={
"label_present": Noul(
instructions=(
"在 `parcel_photo` 中的包裹上是否可见地附有运输标签？"
"寻找地址或配送标签，而不是仅仅制造商的标志。"
),
),
"label_location": Box(
instructions=(
"定位 `parcel_photo` 中附着在包裹上的最大可见运输标签。"
"包含整个标签。"
),
),
},
)
present = response.nouls["label_present"].noul
location = response.boxes["label_location"]
if present < 0.90:
print("未检测到足够确定的运输标签")
elif location.confidence is None or location.confidence < 0.80:
print("标签位置需要人工审查")
else:
edges = location.box
width, height = photo.size
# 添加小边距以减少裁剪标签的可能性。
padding = 0.02
bounds = (
floor(max(0, edges.left - padding) * width),
floor(max(0, edges.top - padding) * height),
ceil(min(1, edges.right + padding) * width),
ceil(min(1, edges.bottom + padding) * height),
)
output = Path(r"d:\images\jev-test\shipping_label_crop.jpg")
photo.crop(bounds).save(output, quality=95)
print("已保存:", output.resolve())
print("框置信度:", location.confidence)

from pathlib import Path
from PIL import ImageDraw
from IPython.display import display
from typesafe_sdk import Image as JevImage, Noul, Box
folder = Path(r"d:\images\jev-test")
# 保留准备好的图像，以便之后在其上绘制。
comparison_photo = load_photo(folder / "car-door-3.png")
response = client.system_one(
state={
"original": image_asset(folder / "car-door-1.png"),
"comparison": JevImage.from_pil(
comparison_photo,
format="jpeg",
),
},
questions={
"visible_difference": Noul(
instructions=(
"比较 `original` 和 `comparison`。 "
"物体、标记、文本、形状或排列是否存在可见差异？查找细微变化，"
"包括添加、删除或修改的细节。 "
"忽略仅由光照、压缩或轻微相机移动引起的差异。"
),
),
"scratch_present": Noul(
instructions=(
"`comparison` 中的汽车门漆面是否显示了 `original` 中不存在的划痕？ "
"区分划痕与反光和面板接缝。"
),
),
"scratch_location": Box(
instructions=(
"在 `comparison` 中定位汽车门漆面中 `original` 中不存在的划痕。 "
"返回一个紧密包围整个划痕的矩形框。 "
"在 `comparison` 中定位，而非 `original`。"
),
),
},
)
probability = response.nouls["visible_difference"].noul
scratch_probability = response.nouls["scratch_present"].noul
location = response.boxes["scratch_location"]
print(f"可见差异的概率: {probability:.3f}")
print(f"存在划痕的概率: {scratch_probability:.3f}")
print("框置信度:", location.confidence)
# 示例阈值：请在自己的图像对上测试这些值。
if probability >= 0.85:
print("很可能存在可见差异")
elif probability <= 0.2:
print("未检测到可见差异")
else:
print("不确定：请手动检查图像")
# Box 总是返回坐标，即使目标不存在。
# 仅当划痕及其位置足够确定时才绘制。
if (
scratch_probability >= 0.8
and location.confidence is not None
and location.confidence >= 0.70
):
edges = location.box
width, height = comparison_photo.size
# 将归一化坐标转换为像素，添加小边距。
padding = 6
rectangle = (
max(0, int(edges.left * width) - padding),
max(0, int(edges.top * height) - padding),
min(width - 1, int(edges.right * width) + padding),
min(height - 1, int(edges.bottom * height) + padding),
)
annotated = comparison_photo.copy()
draw = ImageDraw.Draw(annotated)
draw.rectangle(rectangle, outline="red", width=3)
output_path = folder / "car-door-3-boxed.png"
annotated.save(output_path)
print(f"带注释的图像已保存至: {output_path}")
display(annotated)
else:
print("划痕检测或定位不确定；未绘制框。")

这是我的输出结果。

根据图像差异程度，可能需要调整各种概率阈值以获得所需的输出效果。

## 总结

JEV的图像功能为我们提供了一种实用方法，可以将视觉评估转化为软件可执行的决策。我们可以将图像与参考照片进行分类比较，评估特定质量，检查某物是否存在，或使用新的Box原语定位它。

我对“找不同”这个例子印象特别深刻。虽然年纪渐长，但那个划痕对我来说还挺难发现的——即使我知道它在哪里！

我的四个例子只是展示了JEV图像处理所能提供的机会的一小部分。类似的方法还可以支持设备检测、零售货架监控、农作物损伤评估和制造质量检查等应用场景。

重要的是，应用程序仍然控制后续操作。概率或置信度值可以决定是接受图像、请求重新拍摄还是发送案例进行人工检查。

对感兴趣的朋友，这里是我在文章开头提到的关于OpenAI Decisions API的上一篇TDS文章链接：
 /think