duxingchen/KCGL
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

版本变更V3.35将图像的处理统一更换到新表当中

Browse Source
This commit is contained in:
dxc
2026-05-26 11:28:26 +08:00
parent 682139bab8
commit fb5b8d873b
9 changed files with 281 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
docker-compose.yml
View File

@ -1,4 +1,3 @@
version: '3.8'
services:
db:

14
inventory-backend/.dockerignore
View File

@ -1,6 +1,12 @@
venv/
__pycache__/
.git
.idea
__pycache__
*.pyc
.git/
*.pyo
venv
.venv
env
uploads
pgdata
.env
pgdata/
simhei.ttf

83
inventory-backend/app/api/v1/common/image_search.py
View File

@ -19,6 +19,14 @@ from app.models.base import MaterialBase
# 注册蓝图
image_search_bp = Blueprint('image_search', __name__)
# ============================================================================
# 可配置参数
# ============================================================================
# 以图搜图相似度阈值：余弦距离必须小于此值（距离越小越相似）
# 即余弦相似度 = 1 - 距离，必须 > (1 - SIMILARITY_THRESHOLD)
# 默认 0.25 对应余弦相似度 > 0.75
SIMILARITY_DISTANCE_THRESHOLD = 0.40
# ============================================================================
# POST /api/v1/common/image-search
@ -87,27 +95,80 @@ def image_search():
ie.module_name,
ie.target_id,
ie.image_url,
(1 - (ie.embedding <=> :query_vector)) AS similarity
(1 - (ie.embedding <=> :query_vector)) AS similarity,
(ie.embedding <=> :query_vector) AS distance
FROM image_embeddings ie
WHERE ie.embedding IS NOT NULL
AND (ie.embedding <=> :query_vector) < :distance_threshold
ORDER BY ie.embedding <=> :query_vector
LIMIT 200
""")
raw_records = db.session.execute(sql, {"query_vector": query_vector_str}).fetchall()
raw_records = db.session.execute(sql, {
"query_vector": query_vector_str,
"distance_threshold": SIMILARITY_DISTANCE_THRESHOLD
}).fetchall()
if not raw_records:
return jsonify({"code": 200, "data": []})
return jsonify({"code": 200, "data": [], "msg": "未找到相似图片（阈值过滤后）"})
# 按 (module_name, target_id) 去重，每业务记录只保留最相似的那张图
seen = {}
# ---------------------------------------------------------
# Step 1: 初步去重（同入库单只保留最相似的图片）
# ---------------------------------------------------------
first_img_seen = {}
unique_records = []
for row in raw_records:
key = (row.module_name, row.target_id)
if key not in seen:
seen[key] = row
if key not in first_img_seen:
first_img_seen[key] = True
unique_records.append(row)
# 批量回填业务数据
# ---------------------------------------------------------
# Step 2: 按物料维度去重（相同物料只保留第一条 = 相似度最高的那条）
# ---------------------------------------------------------
target_ids_by_module = {}
for row in seen.values():
for row in unique_records:
target_ids_by_module.setdefault(row.module_name, []).append(row.target_id)
# 查询每条记录的 base_id（跨 stock_buy/semi/product/material_base）
base_id_map = {}
for module in ('stock_buy', 'stock_semi', 'stock_product'):
if module not in target_ids_by_module:
continue
ids = target_ids_by_module[module]
ModelCls = StockBuy if module == 'stock_buy' else (StockSemi if module == 'stock_semi' else StockProduct)
id_col = getattr(ModelCls, 'id')
base_col = getattr(ModelCls, 'base_id')
rows = (
db.session.query(id_col, base_col)
.outerjoin(MaterialBase, base_col == MaterialBase.id)
.filter(id_col.in_(ids))
.all()
)
for rec_id, base_id in rows:
base_id_map[(module, rec_id)] = base_id
if 'material_base' in target_ids_by_module:
for rec_id in target_ids_by_module['material_base']:
base_id_map[('material_base', rec_id)] = rec_id
# 按 base_id 去重：相同物料只保留第一张图
material_seen = {}
final_records = []
for row in unique_records:
base_id = base_id_map.get((row.module_name, row.target_id))
if base_id is not None and base_id in material_seen:
continue
if base_id is not None:
material_seen[base_id] = True
final_records.append(row)
# ---------------------------------------------------------
# Step 3: 批量回填业务数据（基于去重后的 final_records）
# ---------------------------------------------------------
target_ids_by_module = {}
for row in final_records:
target_ids_by_module.setdefault(row.module_name, []).append(row.target_id)
business_map = {}
@ -205,9 +266,9 @@ def image_search():
'url': '/material/index',
}
# 组装最终返回
# 组装最终返回（基于 final_records，按相似度从高到低）
results = []
for row in seen.values():
for row in final_records:
key = (row.module_name, row.target_id)
biz = business_map.get(key, {})
raw_url = row.image_url or ''

30
inventory-backend/app/services/inbound/base_service.py
View File

@ -561,13 +561,20 @@ class MaterialBaseService:
db.session.add(new_material)
db.session.flush() # 获取 new_material.id
# 提取产品图向量到独立表（失败不影响业务）
# 先提交主事务，图片向量异步后台提取
db.session.commit()
image_list = data.get('generalImage', [])
if isinstance(image_list, list) and image_list:
ImageEmbeddingService.save_embeddings(
ImageEmbeddingService.MODULE_MATERIAL_BASE, new_material.id, image_list
from flask import current_app
from app.utils.executor import run_embedding_task
run_embedding_task(
ImageEmbeddingService.save_embeddings_background,
current_app._get_current_object(),
ImageEmbeddingService.MODULE_MATERIAL_BASE,
new_material.id,
image_list
)
db.session.commit()
return new_material
except Exception as e:
@ -597,10 +604,17 @@ class MaterialBaseService:
if 'generalImage' in data:
new_photo_list = data['generalImage']
material.product_image = json.dumps(new_photo_list)
# 保存向量到独立表（全量替换）
ImageEmbeddingService.save_embeddings(
ImageEmbeddingService.MODULE_MATERIAL_BASE, material.id, new_photo_list
)
# 立即触发异步向量提取，不阻塞主事务提交
if isinstance(new_photo_list, list) and new_photo_list:
from flask import current_app
from app.utils.executor import run_embedding_task
run_embedding_task(
ImageEmbeddingService.save_embeddings_background,
current_app._get_current_object(),
ImageEmbeddingService.MODULE_MATERIAL_BASE,
material.id,
new_photo_list
)
else:
material.product_image = None
ImageEmbeddingService.delete_embeddings(

30
inventory-backend/app/services/inbound/buy_service.py
View File

@ -183,13 +183,20 @@ class BuyInboundService:
db.session.add(new_stock)
db.session.flush() # 获取 new_stock.id
# 提取到货图片向量到新表（失败不影响业务）
# 先提交主事务（入库单必须落盘），图片向量异步后台提取
db.session.commit()
photo_list = data.get('arrival_photo', [])
if isinstance(photo_list, list) and photo_list:
ImageEmbeddingService.save_embeddings(
ImageEmbeddingService.MODULE_STOCK_BUY, new_stock.id, photo_list
from flask import current_app
from app.utils.executor import run_embedding_task
run_embedding_task(
ImageEmbeddingService.save_embeddings_background,
current_app._get_current_object(),
ImageEmbeddingService.MODULE_STOCK_BUY,
new_stock.id,
photo_list
)
db.session.commit()
return new_stock
except Exception as e:
db.session.rollback()
@ -254,10 +261,17 @@ class BuyInboundService:
if 'arrival_photo' in data:
new_photo_list = data['arrival_photo']
stock.arrival_photo = json.dumps(new_photo_list)
# 保存向量到独立表（全量替换）
ImageEmbeddingService.save_embeddings(
ImageEmbeddingService.MODULE_STOCK_BUY, stock.id, new_photo_list
)
# 立即触发异步向量提取，不阻塞主事务提交
if isinstance(new_photo_list, list) and new_photo_list:
from flask import current_app
from app.utils.executor import run_embedding_task
run_embedding_task(
ImageEmbeddingService.save_embeddings_background,
current_app._get_current_object(),
ImageEmbeddingService.MODULE_STOCK_BUY,
stock.id,
new_photo_list
)
else:
stock.arrival_photo = None
ImageEmbeddingService.delete_embeddings(

32
inventory-backend/app/services/inbound/product_service.py
View File

@ -189,12 +189,19 @@ class ProductInboundService:
db.session.add(new_stock)
db.session.flush() # 获取 new_stock.id
# 提取产品图片向量到独立表（失败不影响业务）
if isinstance(photo_list, list) and photo_list:
ImageEmbeddingService.save_embeddings(
ImageEmbeddingService.MODULE_STOCK_PRODUCT, new_stock.id, photo_list
)
# 先提交主事务，图片向量异步后台提取
db.session.commit()
if isinstance(photo_list, list) and photo_list:
from flask import current_app
from app.utils.executor import run_embedding_task
run_embedding_task(
ImageEmbeddingService.save_embeddings_background,
current_app._get_current_object(),
ImageEmbeddingService.MODULE_STOCK_PRODUCT,
new_stock.id,
photo_list
)
return new_stock
except Exception as e:
db.session.rollback()
@ -225,10 +232,17 @@ class ProductInboundService:
if 'product_photo' in data:
new_photo_list = data['product_photo']
stock.product_photo = json.dumps(new_photo_list)
# 保存向量到独立表（全量替换）
ImageEmbeddingService.save_embeddings(
ImageEmbeddingService.MODULE_STOCK_PRODUCT, stock.id, new_photo_list
)
# 立即触发异步向量提取，不阻塞主事务提交
if isinstance(new_photo_list, list) and new_photo_list:
from flask import current_app
from app.utils.executor import run_embedding_task
run_embedding_task(
ImageEmbeddingService.save_embeddings_background,
current_app._get_current_object(),
ImageEmbeddingService.MODULE_STOCK_PRODUCT,
stock.id,
new_photo_list
)
else:
stock.product_photo = None
ImageEmbeddingService.delete_embeddings(

32
inventory-backend/app/services/inbound/semi_service.py
View File

@ -226,12 +226,19 @@ class SemiInboundService:
db.session.add(new_stock)
db.session.flush() # 获取 new_stock.id
# 提取到货图片向量到独立表（失败不影响业务）
if isinstance(arrival_list, list) and arrival_list:
ImageEmbeddingService.save_embeddings(
ImageEmbeddingService.MODULE_STOCK_SEMI, new_stock.id, arrival_list
)
# 先提交主事务，图片向量异步后台提取
db.session.commit()
if isinstance(arrival_list, list) and arrival_list:
from flask import current_app
from app.utils.executor import run_embedding_task
run_embedding_task(
ImageEmbeddingService.save_embeddings_background,
current_app._get_current_object(),
ImageEmbeddingService.MODULE_STOCK_SEMI,
new_stock.id,
arrival_list
)
return new_stock
except Exception as e:
db.session.rollback()
@ -280,10 +287,17 @@ class SemiInboundService:
if 'arrival_photo' in data:
new_photo_list = data['arrival_photo']
stock.arrival_photo = json.dumps(new_photo_list)
# 保存向量到独立表（全量替换）
ImageEmbeddingService.save_embeddings(
ImageEmbeddingService.MODULE_STOCK_SEMI, stock.id, new_photo_list
)
# 立即触发异步向量提取，不阻塞主事务提交
if isinstance(new_photo_list, list) and new_photo_list:
from flask import current_app
from app.utils.executor import run_embedding_task
run_embedding_task(
ImageEmbeddingService.save_embeddings_background,
current_app._get_current_object(),
ImageEmbeddingService.MODULE_STOCK_SEMI,
stock.id,
new_photo_list
)
else:
stock.arrival_photo = None
ImageEmbeddingService.delete_embeddings(

127
inventory-backend/app/utils/ai_vision.py
View File

@ -9,6 +9,7 @@ import time
import numpy as np
from PIL import Image
import onnxruntime as ort
import cv2
# ============================================================================
# 全局模型单例（项目启动时加载一次）
@ -48,33 +49,115 @@ IMAGENET_STD = [0.229, 0.224, 0.225]
# 模型输入尺寸
INPUT_SIZE = 224
# ============================================================================
# 背景去除配置：HSV 色彩空间阈值
# ============================================================================
# OpenCV HSV: H∈[0,180], S∈[0,255], V∈[0,255]
# 注意：OpenCV 中 H 通道范围是 0-180（是 OpenCV 自己的标准，和美术的 0-360 对应）
def _center_crop_and_resize(image: Image.Image) -> Image.Image:
# 绿色背景阈值（工业绿幕常用色）
# H: 35~85 对应绿色谱（浅绿到深绿）
# S: 低饱和度（35）到高饱和度（255）
# V: 明暗均可（30~255）
BG_GREEN_LOWER = np.array([35, 35, 30])
BG_GREEN_UPPER = np.array([90, 255, 255])
# 白色/浅色背景阈值（高明度、低饱和度区域）
# H: 不限制（0~180），只看 S 和 V
# S: 很低的饱和度（0~35）→ 接近纯灰/白色
# V: 高明度（180~255）
BG_WHITE_LOWER = np.array([0, 0, 180])
BG_WHITE_UPPER = np.array([180, 40, 255])
# 中性灰填充色（BGR → 转换后 RGB 也是 128,128,128）
NEUTRAL_GRAY_BGR = (128, 128, 128)
def _remove_background(image: Image.Image) -> Image.Image:
"""
CLIP 官方预处理：中心裁剪抗干扰
- 将图片最短边缩放到 224
- 从正中间切取 224x224 区域
利用 OpenCV HSV 色彩空间识别并替换背景为中性灰
支持两种背景类型：
1. 工业绿幕/绿色背景（H: 35~90）
2. 白色/浅色背景（高亮度、低饱和度）
逻辑：
- 将 PIL Image 转为 OpenCV 格式 (RGB → BGR)
- 转 HSV，分别生成绿色掩码和白色掩码
- 合并掩码后，按掩码将背景区域替换为中性灰
- 还原为 PIL Image (BGR → RGB) 返回
参数:
image: PIL Image (RGB, uint8)
返回:
处理后的 PIL Image (RGB, uint8)
"""
# PIL (RGB) → OpenCV (BGR)
img_bgr = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
# 转入 HSV 色彩空间
hsv = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2HSV)
# 生成掩码 1：绿色背景
mask_green = cv2.inRange(hsv, BG_GREEN_LOWER, BG_GREEN_UPPER)
# 生成掩码 2：白色/浅色背景
mask_white = cv2.inRange(hsv, BG_WHITE_LOWER, BG_WHITE_UPPER)
# 合并掩码（任意一种背景都替换）
mask_combined = cv2.bitwise_or(mask_green, mask_white)
# 形态学处理：消除噪点（小面积背景噪点填平）
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
mask_combined = cv2.morphologyEx(mask_combined, cv2.MORPH_CLOSE, kernel)
mask_combined = cv2.morphologyEx(mask_combined, cv2.MORPH_OPEN, kernel)
# 背景替换：将掩码区域填充为中性灰
# 其中 mask_combined=255 的区域为背景，替换为 NEUTRAL_GRAY_BGR
img_bgr_no_bg = img_bgr.copy()
img_bgr_no_bg[mask_combined > 0] = NEUTRAL_GRAY_BGR
# OpenCV (BGR) → PIL (RGB)
result = Image.fromarray(cv2.cvtColor(img_bgr_no_bg, cv2.COLOR_BGR2RGB))
return result
def _letterbox_image(image: Image.Image, size: int = 224) -> Image.Image:
"""
Letterbox 预处理：等比例缩放 + 灰色填充，保持内容不变形
- 将原图最长边缩放到 224
- 短边按相同比例缩放
- 不足部分用 RGB(128,128,128) 灰色填充至 224x224
参数:
image: PIL Image 对象
size: 目标尺寸，默认 224
返回:
224x224 PIL Image
"""
w, h = image.size
# 计算缩放后的目标尺寸
if w < h:
new_w = INPUT_SIZE
new_h = int(h * INPUT_SIZE / w)
else:
new_h = INPUT_SIZE
new_w = int(w * INPUT_SIZE / h)
# 计算缩放比例，使最长边等于 size
scale = size / max(w, h)
new_w = int(w * scale)
new_h = int(h * scale)
# 缩放
image = image.resize((new_w, new_h), Image.BILINEAR)
# 等比例缩放
resized = image.resize((new_w, new_h), Image.LANCZOS)
# 中心裁剪
left = (new_w - INPUT_SIZE) // 2
top = (new_h - INPUT_SIZE) // 2
right = left + INPUT_SIZE
bottom = top + INPUT_SIZE
# 创建灰色画布
canvas = Image.new('RGB', (size, size), (128, 128, 128))
return image.crop((left, top, right, bottom))
# 将缩放后的图片粘贴到画布正中央
paste_x = (size - new_w) // 2
paste_y = (size - new_h) // 2
canvas.paste(resized, (paste_x, paste_y))
return canvas
def _normalize(image_np: np.ndarray) -> np.ndarray:
@ -111,8 +194,12 @@ def get_image_embedding(image_path: str) -> list:
load_clip_model()
# 1. 图片预处理
# Step 1: 背景去除（HSV 色彩空间，绿色/白色背景 → 中性灰替换）
image = Image.open(image_path).convert('RGB')
image = _center_crop_and_resize(image)
image = _remove_background(image)
# Step 2: Letterbox 等比例缩放（保持内容不变形）
image = _letterbox_image(image, INPUT_SIZE)
input_data = _normalize(np.array(image))
input_data = np.expand_dims(input_data, axis=0) # [1, 3, 224, 224]

2
inventory-backend/requirements.txt
View File

@ -10,6 +10,8 @@ flask-cors==4.0.0
redis==5.0.1
# 图片处理核心库
Pillow>=10.0.0
# OpenCV（背景去除、HSV色彩空间抠图）
opencv-python-headless>=4.8.0
# ONNX 模型本地 CPU 推理
onnxruntime>=1.16.0
# 数值计算（ONNX 推理依赖）