【工业计量优选｜水表刻度CRNN文字识别数据集｜YOLO目标检测+PP-OCR双标注｜1500张实拍图像全套训练方案

关注

原创 2026-08-11 11:23:52 243 阅读 AI 写同款· GEO 优化 ›

YOLO在工业缺陷检测中的落地实践本文系统阐述YOLO在工业缺陷检测中的落地实践，针对产线节拍快、缺陷微小、样本不平衡等挑战，提出从成像优化、数据构建到模型部署的全流程方法论。强调缺陷定义标准化、多模态融合（2D+3D）、混合架构（传统视觉+深度学习）与可追溯工程化设计，实现高精度、低误判、稳定可靠的产线应用，为工业视觉项目提供可复用的技术框架。阅读详情

在这里插入图片描述
crnn的水表刻度识别
数据集包含1类别
收集整理标注好的数据共1500张
包含yolo格式数据集+ppocr识别格式数据集+训练好的模型
支持数据转换voc

一、数据集概述

面向水务智能抄表、园区远程水表读数自动化采集、工业仪表视觉读取场景研发，整套方案采用两级串联识别架构：先用YOLOv8定位水表表盘区域，裁剪表盘后送入CRNN卷积循环神经网络完成数字刻度序列识别；数据集共计1500张现场实拍水表图像，单类别表盘目标，同步搭载YOLO检测标注、PP-OCR文本标注两套标签，附带VOC格式转换脚本、训练权重、完整工程代码，适配水务智能化课题研发与落地部署。

二、核心信息概览

维度	详细参数
识别架构	YOLOv8表盘定位 + CRNN数字序列文字识别（两级流水线）
目标类别	1类：water_meter（水表表盘）
图像总量	1500张实拍水表图像
标注体系	YOLO txt检测标注 + PP-OCR json文本标注，附带VOC转换工具
适配分辨率	统一640×640
配套资源	数据集划分脚本、YOLOv8训练推理代码、CRNN网络完整源码、预训练权重、临时文件清理工具
适配场景	水务远程抄表、园区仪表巡检、工业数显仪表字符识别

三、数据集核心亮点

✅ 双任务配套标注：检测框+刻度文本一一对应，完美适配先定位后识别流水线方案
✅ 实拍复杂工况数据：涵盖强光、阴影、轻微污渍等真实水表使用环境，模型泛化性更强
✅ 一站式完整链路：数据集划分→表盘检测→裁剪→CRNN序列识别全流程代码无需二次开发
✅ 多格式兼容：原生YOLO、PP-OCR标注，附赠VOC转换脚本，兼容Faster R-CNN等两阶段算法
✅ 轻量化部署友好：YOLOv8n+小型CRNN组合，可移植嵌入式设备做现场抄表识别
✅ 附带训练好模型权重，可直接开箱推理测试效果

四、适用场景

水务公司远程自动抄表系统、小区物业水表集中巡检识别
工业数显仪表、流量计、压力表字符序列视觉读取课题研究
YOLO目标检测+CRNN文本识别串联流水线算法实验
嵌入式边缘端水表识别项目开发、机器视觉课程实训
水务智能化、物联网视觉采集相关科研课题

五、整体项目目录结构

water_meter_recognition/
├── datasets/
│   └── water_meter_dataset/
│       ├── images/        # 原始水表图片
│       ├── labels/        # YOLO检测txt标注
│       ├── ocr_labels/    # PP-OCR json文本标注
│       ├── train/         # 划分后训练集
│       │   ├── images/
│       │   ├── labels/
│       │   └── ocr_labels/
│       └── val/           # 划分后验证集
│           ├── images/
│           ├── labels/
│           └── ocr_labels/
├── runs/detect/            # YOLO训练日志与权重
├── weights/
│   ├── best.pt             # YOLO表盘检测最优权重
│   └── crnn_water_meter.pth # CRNN数字识别权重
├── data.yaml               # YOLO数据集配置文件
└── main.py                 # 整套流程整合脚本

六、分步完整实操流程

步骤1：数据集划分（训练集8:验证集2）

import os
import random
from pathlib import Path
import shutil

def split_dataset(data_dir, train_ratio=0.8):
    images = list(Path(data_dir).glob('*.jpg'))
    random.seed(42)
    random.shuffle(images)
    
    num_train = int(len(images) * train_ratio)
    train_images = images[:num_train]
    val_images = images[num_train:]
    
    train_dir = Path(data_dir).parent / 'train'
    val_dir = Path(data_dir).parent / 'val'
    
    train_img_dir = train_dir / 'images'
    train_label_dir = train_dir / 'labels'
    train_ocr_label_dir = train_dir / 'ocr_labels'
    
    val_img_dir = val_dir / 'images'
    val_label_dir = val_dir / 'labels'
    val_ocr_label_dir = val_dir / 'ocr_labels'
    
    for folder in [train_img_dir,train_label_dir,train_ocr_label_dir,val_img_dir,val_label_dir,val_ocr_label_dir]:
        folder.mkdir(parents=True, exist_ok=True)
    
    for img in train_images:
        label_path = img.with_suffix('.txt')
        ocr_label_path = img.with_name(img.stem + '.json')
        shutil.copy(img, train_img_dir / img.name)
        shutil.copy(label_path, train_label_dir / label_path.name)
        shutil.copy(ocr_label_path, train_ocr_label_dir / ocr_label_path.name)
    
    for img in val_images:
        label_path = img.with_suffix('.txt')
        ocr_label_path = img.with_name(img.stem + '.json')
        shutil.copy(img, val_img_dir / img.name)
        shutil.copy(label_path, val_label_dir / label_path.name)
        shutil.copy(ocr_label_path, val_ocr_label_dir / ocr_label_path.name)

split_dataset('./datasets/water_meter_dataset/images')

步骤2：YOLO数据集配置 data.yaml

train: ./datasets/water_meter_dataset/train/images
val: ./datasets/water_meter_dataset/val/images
nc: 1
names: ['water_meter']

步骤3：YOLOv8表盘区域检测训练+评估+推理

import torch
from ultralytics import YOLO

torch.manual_seed(42)
model = YOLO('yolov8n.pt')

# 训练
results = model.train(
    data="data.yaml",
    epochs=100,
    imgsz=640,
    batch=16,
    name='water_meter_detection',
    project='runs/detect'
)
# 验证集评估
metrics = model.val()
best_weights = 'runs/detect/water_meter_detection/weights/best.pt'
print("最优检测权重保存路径：", best_weights)

步骤4：CRNN卷积循环网络数字刻度识别训练

CRNN核心逻辑：灰度表盘图像经卷积提取空间特征，LSTM时序建模数字前后序列依赖，完成水表连续读数识别

import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import transforms
from torch.utils.data import DataLoader, Dataset
import json
import cv2
import numpy as np

class WaterMeterOCRDataset(Dataset):
    def __init__(self, root_dir, transform=None):
        self.root = root_dir
        self.transform = transform
        self.img_list = sorted(os.listdir(os.path.join(root_dir,"images")))
        self.ocr_list = sorted(os.listdir(os.path.join(root_dir,"ocr_labels")))
    def __len__(self):
        return len(self.img_list)
    def __getitem__(self, idx):
        img_path = os.path.join(self.root,"images",self.img_list[idx])
        json_path = os.path.join(self.root,"ocr_labels",self.ocr_list[idx])
        img = cv2.imread(img_path)
        img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
        img = np.expand_dims(img, -1) / 255.0
        with open(json_path,"r",encoding="utf-8") as f:
            text = json.load(f)["text"]
        if self.transform:
            img = self.transform(img)
        return img, text

# CRNN网络结构
class CRNN(nn.Module):
    def __init__(self, in_feat=160, hidden=128, out=10, layers=2):
        super().__init__()
        self.conv = nn.Sequential(
            nn.Conv2d(1,32,3,padding=1),nn.ReLU(),nn.MaxPool2d(2,2),
            nn.Conv2d(32,64,3,padding=1),nn.ReLU(),nn.MaxPool2d(2,2)
        )
        self.lstm = nn.LSTM(in_feat, hidden, layers, bidirectional=True)
        self.fc = nn.Linear(hidden*2, out)
    def forward(self, x):
        x = self.conv(x)
        x = x.squeeze(2).permute(2,0,1)
        out,_ = self.lstm(x)
        return self.fc(out[-1])

# 训练参数
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
trans = transforms.Compose([transforms.ToTensor(),transforms.Resize((64,160))])
train_set = WaterMeterOCRDataset("./datasets/water_meter_dataset/train", trans)
val_set = WaterMeterOCRDataset("./datasets/water_meter_dataset/val", trans)
train_loader = DataLoader(train_set,16,shuffle=True)
val_loader = DataLoader(val_set,16,False)

crnn = CRNN().to(device)
opt = optim.Adam(crnn.parameters(),lr=1e-3)
loss_fn = nn.CrossEntropyLoss()

# 迭代训练
for epoch in range(10):
    crnn.train()
    total_loss = 0
    for imgs, texts in train_loader:
        imgs = imgs.to(device)
        targets = torch.tensor([int(t) for t in texts]).to(device)
        opt.zero_grad()
        pred = crnn(imgs)
        loss = loss_fn(pred, targets)
        loss.backward()
        opt.step()
        total_loss += loss.item()
    print(f"Epoch{epoch+1} Loss:{total_loss/len(train_loader):.4f}")
torch.save(crnn.state_dict(),"weights/crnn_water_meter.pth")

# 验证集准确率评估
crnn.eval()
correct, total = 0,0
with torch.no_grad():
    for imgs, texts in val_loader:
        imgs = imgs.to(device)
        targets = torch.tensor([int(t) for t in texts]).to(device)
        pred = crnn(imgs)
        _,res = torch.max(pred,1)
        total += targets.size(0)
        correct += (res==targets).sum().item()
print(f"验证集识别准确率：{100*correct/total:.2f}%")

步骤5：整套流水线联合推理（检测表盘+裁剪+CRNN读数）

import cv2
from PIL import Image,ImageDraw
from ultralytics import YOLO

det_model = YOLO("weights/best.pt")
ocr_model = CRNN().to(device)
ocr_model.load_state_dict(torch.load("weights/crnn_water_meter.pth"))
ocr_model.eval()

def meter_full_infer(img_path):
    # 1. YOLO检测表盘
    res = det_model(img_path,conf=0.5)[0]
    box = res.boxes.xyxy[0].cpu().numpy().astype(int)
    x1,y1,x2,y2 = box
    # 2. 裁剪表盘区域
    raw_img = cv2.imread(img_path)
    crop_img = raw_img[y1:y2,x1:x2]
    # 3. CRNN识别数字
    gray = cv2.cvtColor(crop_img,cv2.COLOR_BGR2GRAY)
    gray = cv2.resize(gray,(160,64))
    tensor = trans(np.expand_dims(gray,-1)).unsqueeze(0).to(device)
    with torch.no_grad():
        out = ocr_model(tensor)
        num = torch.argmax(out).item()
    # 4. 绘制结果
    pil_img = Image.open(img_path)
    draw = ImageDraw.Draw(pil_img)
    draw.rectangle([x1,y1,x2,y2],outline="green",width=3)
    draw.text((x1,y1-25),f"水表读数：{num}",fill="red")
    pil_img.save("meter_result.jpg")
    pil_img.show()
    return num

meter_full_infer("test.jpg")

步骤6：临时缓存文件清理

import shutil,os
def clean_cache(path):
    for p in [f"{path}/wandb",f"{path}/cache"]:
        if os.path.exists(p):
            shutil.rmtree(p)
clean_cache("runs/detect/water_meter_detection")