YOLO26 识别验证码
最近 YOLO26 出来了，我还没使用过 yolo ,用它来做一个测试。注意：直接用OCR 识别验证码更加方便。用YOLO 不太合适，我只是为了做测试！
先看效果！
codes.yml 文件如下：
# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license # COCO8 dataset (first 8 images from COCO train2017) by Ultralytics # Documentation: https://docs.ultralytics.com/datasets/detect/coco8/ # Example usage: yolo train data=coco8.yaml # parent # ├── ultralytics # └── datasets # └── coco8 ← downloads here (1 MB) # Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] path: D:\learn\yolo\dataset\yolo_dataset_2026-04-28 # dataset root dir train: images/train # train images (relative to 'path') 4 images val: images/val # val images (relative to 'path') 4 images test: # test images (optional) # Classes names: 0: "0" 1: "1" 2: "2" 3: "3" 4: "4" 5: "5" 6: "6" 7: "7" 8: "8" 9: "9" 10: "+" 11: "-" 12: "*" 13: "/" 14: "=" 15: "?"
yolo cli 命令如下
#训练，由于只是测试，所以没配置其他训练参数 yolo detect train data="D:\learn\yolo\ultralytics\ultralytics\cfg\datasets\codes.yaml" model=yolo26n.pt epochs=100 imgsz=640 batch =0.5 workers=1 #验证 yolo detect val model=codes.pt data="D:\learn\yolo\ultralytics\ultralytics\cfg\datasets\codes.yaml" nc=16 #预测 yolo predict model=codes.pt source="D:\learn\yolo\ultralytics\ultralytics\test\codes_processed.png" #导出 为了 java 代码 https://gitee.com/agricultureiot/yolo-onnx-java 导出成 onnx，这个项目只支持 opset 19 以下
yolo export model=codes.pt format=onnx opset=18
python 脚本如下：
import time
import requests
from ultralytics import YOLO
from PIL import Image
import numpy as np
# Load a pretrained YOLO26n model
model = YOLO("../codes2.pt")
# 生成随机码并下载验证码图片
random_str = str(int(time.time() * 1000))
# java pig4cloud 管理后台验证码地址
url = f"https://www.xxxx.cn/auth/code/image?randomStr={random_str}"
print(f"正在下载验证码图片...")
print(f"请求URL: {url}")
try:
response = requests.get(url, timeout=10)
response.raise_for_status()
# 保存图片到上一级目录
with open("codes.png", "wb") as f:
f.write(response.content)
print("验证码图片下载成功: codes.png")
except Exception as e:
print(f"下载验证码图片失败: {e}")
exit(1)
results = model("codes.png")
# 处理预测结果
for result in results:
# 获取检测框
boxes = result.boxes
if boxes is not None:
# 获取分类ID
class_ids = boxes.cls.cpu().numpy()
# 获取置信度
confidences = boxes.conf.cpu().numpy()
# 获取边界框坐标 (xyxy格式: x1, y1, x2, y2)
box_coords = boxes.xyxy.cpu().numpy()
# 获取分类名称列表
names = result.names
# 将检测结果组合成列表，每个元素包含位置信息和检测信息
detections = []
for i, (class_id, confidence, box) in enumerate(zip(class_ids, confidences, box_coords)):
x1, y1, x2, y2 = box
center_x = (x1 + x2) / 2 # 计算中心点x坐标
detections.append({
'index': i,
'class_id': int(class_id),
'confidence': confidence,
'class_name': names[int(class_id)],
'box': box,
'center_x': center_x,
'x1': x1
})
# 按照中心点x坐标从左到右排序
detections.sort(key=lambda d: d['center_x'])
# 打印排序后的检测结果
print(f"检测到 {len(detections)} 个目标（已按从左到右排序）:")
print("-" * 50)
def calculate_overlap_ratio(box1, box2):
"""计算两个框的重叠比例（相对于较小框的面积）"""
x1_1, y1_1, x2_1, y2_1 = box1
x1_2, y1_2, x2_2, y2_2 = box2
# 计算交集区域
inter_x1 = max(x1_1, x1_2)
inter_y1 = max(y1_1, y1_2)
inter_x2 = min(x2_1, x2_2)
inter_y2 = min(y2_1, y2_2)
# 检查是否有交集
if inter_x2 < inter_x1 or inter_y2 < inter_y1:
return 0.0
# 计算交集面积
inter_area = (inter_x2 - inter_x1) * (inter_y2 - inter_y1)
# 计算两个框的面积
area1 = (x2_1 - x1_1) * (y2_1 - y1_1)
area2 = (x2_2 - x1_2) * (y2_2 - y1_2)
# 使用较小框的面积作为基准
min_area = min(area1, area2)
if min_area == 0:
return 0.0
# 返回重叠比例
return inter_area / min_area
def filter_overlapping_detections(detections, overlap_threshold=0.5):
"""过滤重叠的检测框，对于相同class_id且重叠超过阈值的，只保留第一个"""
filtered = []
used_indices = set()
for i in range(len(detections)):
if i in used_indices:
continue
# 添加当前检测到结果列表
filtered.append(i)
# 检查后续检测是否与当前检测重叠
for j in range(i + 1, len(detections)):
if j in used_indices:
continue
# 只检查相同class_id的
if detections[i]['class_id'] != detections[j]['class_id']:
continue
# 计算重叠比例
overlap_ratio = calculate_overlap_ratio(
detections[i]['box'],
detections[j]['box']
)
# 如果重叠超过阈值，标记为已使用（跳过）
if overlap_ratio > overlap_threshold:
used_indices.add(j)
print(f" [过滤] 目标{j + 1}与目标{i + 1}重叠{overlap_ratio:.2%}，已跳过")
return filtered
# 过滤重叠的检测
valid_indices = filter_overlapping_detections(detections, overlap_threshold=0.5)
allClasName = ""
for idx, det_idx in enumerate(valid_indices, 1):
det = detections[det_idx]
print(f"目标 {idx}:")
print(f" 分类ID: {det['class_id']}")
print(f" 分类名称: {det['class_name']}")
print(f" 置信度: {det['confidence']:.4f}")
print(
f" 边界框: x1={det['box'][0]:.1f}, y1={det['box'][1]:.1f}, x2={det['box'][2]:.1f}, y2={det['box'][3]:.1f}")
print(f" 中心点x坐标: {det['center_x']:.1f}")
print("-" * 50)
allClasName = f"{allClasName} {det['class_name']}"
print(allClasName)
else:
print("未检测到任何目标")
不要问我python 什么意思，我也不懂。ai给我生成的！！！
简诉下遇到的问题：
1.训练图片数量比较少，准确率低
2.标注错了好多图片，导致准确率非常低
3.由于数字是空心的，导致部分数字重复识别出来，例如 3+2=？ 的图片识别出来 3 3 + 2 = ？,这个时候 class_id 重复，且堆叠 超过 50% 的，我就丢弃后面重复的结果得到 3 + 2 = ？
浙公网安备 33010602011771号