paddle ocr

[paddle ocr笔记](#paddle ocr笔记)
- 准备工作
- reference
- [to onnx](#to onnx)
- 文本检测
- 文本检测
- 文字识别

paddle ocr笔记

准备工作

下载字典ppocr_keys_v1.txt，下标从1开始
模型转换

reference

paddlepaddle

to onnx

下载模型，或者直接使用python跑一下并且把本地模型拿过来用，一共三个模型：文本检测，文本方向校准，文字识别模型

复制代码

paddle2onnx --model_dir ./det/ch/ch_PP-OCRv4_det_infer --model_filename inference.pdmodel  --params_filename inference.pdiparams  --save_file ./det-model.onnx  --opset_version 11  --enable_onnx_checker True

paddle2onnx --model_dir ./rec/ch/ch_PP-OCRv4_rec_infer --model_filename inference.pdmodel  --params_filename inference.pdiparams  --save_file ./rec-model.onnx  --opset_version 11  --enable_onnx_checker True

paddle2onnx --model_dir ./cls/ch_ppocr_mobile_v2.0_cls_infer --model_filename inference.pdmodel  --params_filename inference.pdiparams  --save_file ./cls-padmodel.onnx  --opset_version 11  --enable_onnx_checker True

文本检测

复制代码

import onnxruntime
import cv2
import numpy as np

# 读取图片
# image = cv2.imread("chinese.png")
image = cv2.imread("none.png")

# 加载 ONNX 文本检测模型
det_session = onnxruntime.InferenceSession("./onnx/det-model.onnx")

# 预处理
det_input = cv2.resize(image, (640, 640)) / 255.0  # 归一化
det_input = np.transpose(det_input, (2, 0, 1))[np.newaxis, :, :, :].astype(np.float32)

# 推理
input_name = det_session.get_inputs()[0].name
output_name = det_session.get_outputs()[0].name
det_output = det_session.run([output_name], {input_name: det_input})[0]

# 获取非零值的索引（文本区域）
non_zero_indices = np.nonzero(det_output)

print(f'--------------------------')
print(f'len:{len(non_zero_indices[0])}')
# 打印非零值的索引及对应的值
for i in range(len(non_zero_indices[0])):
    print(f'--------------------------')
    coords = tuple(non_zero_indices[j][i] for j in range(len(non_zero_indices)))
    print(f"坐标: {coords}, 值: {det_output[coords]}")

文本检测

复制代码

import onnxruntime
import cv2
import numpy as np

# 读取图片
# image = cv2.imread("chinese.png")
# image = cv2.imread("none.png")
image = cv2.imread("oneline.png")

# 加载 ONNX 文本检测模型
det_session = onnxruntime.InferenceSession("./onnx/det-model.onnx")

# 预处理
det_input = cv2.resize(image, (640, 640)) / 255.0  # 归一化
# det_input = image / 255.0  # 归一化
det_input = np.transpose(det_input, (2, 0, 1))[np.newaxis, :, :, :].astype(np.float32)

# 推理
input_name = det_session.get_inputs()[0].name
output_name = det_session.get_outputs()[0].name
det_output = det_session.run([output_name], {input_name: det_input})[0]



#
#
#
# 获取非零值的索引（文本区域）
# non_zero_indices = np.nonzero(det_output)

# print(f'--------------------------')
# print(f'len:{len(non_zero_indices[0])}')
# # 打印非零值的索引及对应的值
# for i in range(len(non_zero_indices[0])):
#     print(f'--------------------------')
#     coords = tuple(non_zero_indices[j][i] for j in range(len(non_zero_indices)))
#     print(f"坐标: {coords}, 值: {det_output[coords]}")

# 5. 确保 `det_output` 是 4D (batch, channels, height, width)
print("det_output shape:", det_output.shape)

# 6. 获取二值化文本区域
box_thresh = 0.3
text_mask = (det_output[0, 0, :, :] > box_thresh).astype(np.uint8) * 255

# 7. 确保 `text_mask` 是单通道 2D 图像
print("text_mask shape:", text_mask.shape)  # 应该是 (height, width)

# 8. 显示二值化结果（可选）
cv2.imshow("Text Mask", text_mask)
cv2.waitKey(0)
cv2.destroyAllWindows()

# 9. 确保 `text_mask` 兼容 OpenCV
if len(text_mask.shape) == 3:
    text_mask = cv2.cvtColor(text_mask, cv2.COLOR_BGR2GRAY)

# 10. 轮廓检测
contours, _ = cv2.findContours(text_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

# 11. 打印检测到的文本框
for contour in contours:
    x, y, w, h = cv2.boundingRect(contour)
    print(f"检测到的文本框: x={x}, y={y}, w={w}, h={h}")

文字识别

复制代码

import onnxruntime
import cv2
import numpy as np

# 读取图片
image = cv2.imread("rec-cv2.imread.png")

target_height = 48
h, w = image.shape[:2]
new_w = int(w * (target_height / h))
new_w = max(32, (new_w // 32) * 32)
image = cv2.resize(image, (new_w, target_height));
print(f'resize: {image.shape}')

# image = cv2.imread("none.png")

# 加载 ONNX 文本检测模型
rec_session = onnxruntime.InferenceSession("./onnx/rec-model.onnx")

# 预处理
# rec_input = cv2.resize(image, (640, 640)) / 255.0  # 归一化
rec_input = image / 255.0  # 归一化
rec_input = np.transpose(rec_input, (2, 0, 1))[np.newaxis, :, :, :].astype(np.float32)

print(f'shape: {rec_input.shape}')

# 推理
input_name = rec_session.get_inputs()[0].name
output_name = rec_session.get_outputs()[0].name
print(f'input_name: {input_name}')
print(f'output_name: {output_name}')
rec_output = rec_session.run([output_name], {input_name: rec_input})[0]

with open('onnx/dict.txt', 'r', encoding='utf-8') as f:
    dict_list = f.readlines()
    dict_list = [line.strip() for line in dict_list]
    # 返回下标的字典
    dict_index = {i + 1: char for i, char in enumerate(dict_list)}

print(rec_output[0].shape)


# 假设 rec_output[0] 的 shape 为 (60, 6625)
probabilities = rec_output[0]  # 取出 (60, 6625)

# Step 1: 找到每个时间步概率最高的索引
max_indices = np.argmax(probabilities, axis=1)  # shape: (60,)

# print(f'--dict: {dict_index}')
# Step 2: CTC 处理（去重）
decoded_text = []
prev_index = -1  # 记录前一个索引，避免重复
for index in max_indices:
    if index != prev_index and index in dict_index:  # 避免重复 & 确保索引合法
        print(f'--index: {index}, dict[{index}] = dict_index[{index}]')
        decoded_text.append(dict_index[index])
    prev_index = index

# Step 3: 转换为字符串
decoded_text = "".join(decoded_text)

print("识别结果:", decoded_text)