haotian/kangda_robotic_dog
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

1.实现使用paddleocr的rknn脚本

Browse Source
This commit is contained in:
haotian 2025-08-26 15:38:04 +08:00
parent fcc1313ac4
commit 1e7e888dcd
1 changed files with 557 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

557
010使用PaddleOCR_rknn.py Normal file
View File

@ -0,0 +1,557 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
RK3588 PaddleOCR RKNN推理程序
使用转换后的RKNN模型在RK3588上进行OCR文本检测和识别
"""
import cv2
import yaml
import numpy as np
import math
from rknn.api import RKNN
import argparse
import os
import time
class RK3588OCR:
def __init__(self, det_model_path, rec_model_path):
"""
初始化OCR推理器
Args:
det_model_path: 文本检测RKNN模型路径
rec_model_path: 文本识别RKNN模型路径
"""
self.det_model_path = det_model_path
self.rec_model_path = rec_model_path
# 初始化RKNN实例
self.det_rknn = RKNN(verbose=False)
self.rec_rknn = RKNN(verbose=False)
# 模型输入尺寸
self.det_input_size = (640, 640)
self.rec_input_size = (320, 48)
# 文本检测相关参数
self.det_threshold = 0.3
self.det_box_threshold = 0.6
self.det_unclip_ratio = 1.5
# 加载模型
self._load_models()
self.character = self.get_dict()
def get_dict(self, dict_path='/home/orangepi/Desktop/kangda_robotic_dog/机器狗后台服务/dict.yaml'):
"""
加载字典
"""
with open(dict_path, 'r', encoding='utf-8') as f:
dict_rec = yaml.safe_load(f)
return dict_rec.get('character_dict', [])
def _load_models(self):
"""加载RKNN模型"""
print("加载文本检测模型...")
ret = self.det_rknn.load_rknn(self.det_model_path)
if ret != 0:
raise Exception(f"加载检测模型失败: {ret}")
ret = self.det_rknn.init_runtime(target='rk3588')
if ret != 0:
raise Exception(f"初始化检测模型运行环境失败: {ret}")
print("加载文本识别模型...")
ret = self.rec_rknn.load_rknn(self.rec_model_path)
if ret != 0:
raise Exception(f"加载识别模型失败: {ret}")
ret = self.rec_rknn.init_runtime(target='rk3588')
if ret != 0:
raise Exception(f"初始化识别模型运行环境失败: {ret}")
print("模型加载完成!")
def resize_norm_img_det(self, img, input_shape=(640, 640)):
"""
检测模型的图像预处理 - 固定输入形状 [1, 3, 640, 640]
"""
h, w, _ = img.shape
target_h, target_w = input_shape
# 计算缩放比例 - 保持宽高比
ratio_h = target_h / h
ratio_w = target_w / w
ratio = min(ratio_h, ratio_w)
# 计算缩放后的尺寸
new_h = int(h * ratio)
new_w = int(w * ratio)
# 调整图像大小
resized_img = cv2.resize(img, (new_w, new_h))
# 创建目标尺寸的图像，用灰色填充
padded_img = np.ones((target_h, target_w, 3), dtype=np.float32) * 114.0 # 直接用float32
# 计算居中位置
top = (target_h - new_h) // 2
left = (target_w - new_w) // 2
# 将缩放后的图像放到居中位置
padded_img[top:top+new_h, left:left+new_w] = resized_img.astype(np.float32)
# 归一化
img = (padded_img / 255.0 - np.array([0.485, 0.456, 0.406], dtype=np.float32)) / np.array([0.229, 0.224, 0.225], dtype=np.float32)
img = img.transpose(2, 0, 1).astype(np.float32)
img = np.expand_dims(img, axis=0).astype(np.float32)
return img, ratio, (top, left)
def post_process_det(self, dt_boxes, ratio, padding_info, ori_shape):
"""
检测结果后处理 - 适配固定输入形状
"""
if dt_boxes is None:
return None
ori_h, ori_w = ori_shape
top, left = padding_info
# 将坐标从模型输出空间转换回原图空间
dt_boxes[:, :, 0] = (dt_boxes[:, :, 0] - left) / ratio
dt_boxes[:, :, 1] = (dt_boxes[:, :, 1] - top) / ratio
# 裁剪到原图范围内
dt_boxes[:, :, 0] = np.clip(dt_boxes[:, :, 0], 0, ori_w)
dt_boxes[:, :, 1] = np.clip(dt_boxes[:, :, 1], 0, ori_h)
return dt_boxes
def boxes_from_bitmap(self, pred, bitmap, dest_width, dest_height, max_candidates=1000, box_thresh=0.6):
"""
从位图中提取文本框
"""
bitmap = bitmap.astype(np.uint8)
height, width = bitmap.shape
# 查找轮廓
contours, _ = cv2.findContours(bitmap, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
num_contours = min(len(contours), max_candidates)
boxes = []
scores = []
for i in range(num_contours):
contour = contours[i]
points, sside = self.get_mini_boxes(contour)
if sside < 5:
continue
points = np.array(points)
score = self.box_score_fast(pred, points.reshape(-1, 2))
if box_thresh > score:
continue
# 扩展box
box = self.unclip(points, 1.5).reshape(-1, 1, 2)
box, sside = self.get_mini_boxes(box)
if sside < 5 + 2:
continue
box = np.array(box)
box[:, 0] = np.clip(box[:, 0] / width * dest_width, 0, dest_width)
box[:, 1] = np.clip(box[:, 1] / height * dest_height, 0, dest_height)
boxes.append(box.astype(np.int16))
scores.append(score)
return np.array(boxes), scores
def get_mini_boxes(self, contour):
"""获取最小外接矩形"""
bounding_box = cv2.minAreaRect(contour)
points = sorted(list(cv2.boxPoints(bounding_box)), key=lambda x: x[0])
index_1, index_2, index_3, index_4 = 0, 1, 2, 3
if points[1][1] > points[0][1]:
index_1 = 0
index_4 = 1
else:
index_1 = 1
index_4 = 0
if points[3][1] > points[2][1]:
index_2 = 2
index_3 = 3
else:
index_2 = 3
index_3 = 2
box = [points[index_1], points[index_2], points[index_3], points[index_4]]
return box, min(bounding_box[1])
def box_score_fast(self, bitmap, _box):
"""快速计算box得分"""
h, w = bitmap.shape[:2]
box = _box.copy()
xmin = np.clip(np.floor(box[:, 0].min()).astype(int), 0, w - 1)
xmax = np.clip(np.ceil(box[:, 0].max()).astype(int), 0, w - 1)
ymin = np.clip(np.floor(box[:, 1].min()).astype(int), 0, h - 1)
ymax = np.clip(np.ceil(box[:, 1].max()).astype(int), 0, h - 1)
mask = np.zeros((ymax - ymin + 1, xmax - xmin + 1), dtype=np.uint8)
box[:, 0] = box[:, 0] - xmin
box[:, 1] = box[:, 1] - ymin
cv2.fillPoly(mask, box.reshape(1, -1, 2).astype(np.int32), 1)
return cv2.mean(bitmap[ymin:ymax + 1, xmin:xmax + 1], mask)[0]
def unclip(self, box, unclip_ratio):
"""扩展文本框"""
from shapely.geometry import Polygon
import pyclipper
poly = Polygon(box)
distance = poly.area * unclip_ratio / poly.length
offset = pyclipper.PyclipperOffset()
offset.AddPath(box, pyclipper.JT_ROUND, pyclipper.ET_CLOSEDPOLYGON)
expanded = offset.Execute(distance)
if len(expanded) == 0:
return box
else:
return np.array(expanded[0])
def resize_norm_img_rec(self, img, input_shape=(320, 48)):
"""
识别模型的图像预处理 - 固定输入形状 [1, 3, 48, 320]
"""
target_w, target_h = input_shape # 注意：宽度在前
h, w = img.shape[:2]
# 计算缩放比例，保持宽高比
ratio_h = target_h / h
ratio_w = target_w / w
ratio = min(ratio_h, ratio_w)
# 计算缩放后的尺寸
new_h = int(h * ratio)
new_w = int(w * ratio)
# 调整图像大小
resized_image = cv2.resize(img, (new_w, new_h))
# 创建目标尺寸的图像，用黑色填充
padded_image = np.zeros((target_h, target_w, 3), dtype=np.float32) # 直接用float32
# 将缩放后的图像放到左上角（识别模型通常左对齐）
padded_image[:new_h, :new_w] = resized_image.astype(np.float32)
# 归一化
# padded_image = (padded_image / 255.0 - np.array([0.485, 0.456, 0.406], dtype=np.float32)) / np.array([0.229, 0.224, 0.225], dtype=np.float32)
# 不缩放反而会将识别结果再移后一个??
padded_image = (padded_image / 255.0).astype(np.float32)
padded_image = padded_image.transpose((2, 0, 1)).astype(np.float32)
return np.expand_dims(padded_image, axis=0).astype(np.float32)
def decode_rec_result(self, preds_prob):
"""
解码识别结果
"""
# preds_idx = preds_idx[0]
preds_prob = preds_prob[0]
preds_idx = np.argmax(preds_prob, axis=1)
preds_prob = np.max(preds_prob, axis=1)
# CTC解码
last_idx = 0
preds_text = []
preds_conf = []
# print("preds_id", len(preds_idx[0]))
for i, idx in enumerate(preds_idx):
if idx != last_idx and idx != 0: # 0是blank
if idx < len(self.character):
# print("self.character[idx]", self.character[idx])
# print("preds_prob[i]", preds_prob[i])
preds_text.append(self.character[idx])
preds_conf.append(preds_prob[i])
last_idx = idx
text = ''.join(preds_text)
conf = np.mean(preds_conf) if preds_conf else 0.0
return text, conf
def detect_text(self, image):
"""
文本检测 - 适配固定输入形状 [1, 3, 640, 640]
"""
ori_h, ori_w = image.shape[:2]
# 预处理
det_img, ratio, padding_info = self.resize_norm_img_det(image)
# 推理
det_output = self.det_rknn.inference(inputs=[det_img], data_format="nchw")[0]
# 后处理
mask = det_output[0, 0, :, :]
threshold = 0.3
bitmap = (mask > threshold).astype(np.uint8) * 255
# 从位图中提取文本框（坐标是在640x640空间中的）
boxes, scores = self.boxes_from_bitmap(mask, bitmap, 640, 640)
# 将坐标转换回原图空间
if len(boxes) > 0:
boxes = self.post_process_det(boxes, ratio, padding_info, (ori_h, ori_w))
print("*"*100, len(boxes))
return boxes, scores
def visualize_det_results(self, image_path, boxes):
image = cv2.imread(image_path)
for box in boxes:
box = np.array(box, dtype=np.int32)
cv2.polylines(image, [box], True, (0, 255, 0), 2)
cv2.imwrite('./visual_det.jpg', image)
def recognize_text(self, image):
"""
文本识别
"""
# 预处理
rec_img = self.resize_norm_img_rec(image)
# 推理
rec_output = self.rec_rknn.inference(inputs=[rec_img], data_format="nchw")
# 解码
text, conf = self.decode_rec_result(rec_output[0])
# print("")
return text, conf
def get_rotate_crop_image(self, img, points):
"""
根据四个点坐标裁剪并矫正图像
"""
img_crop_width = int(
max(
np.linalg.norm(points[0] - points[1]),
np.linalg.norm(points[2] - points[3])))
img_crop_height = int(
max(
np.linalg.norm(points[0] - points[3]),
np.linalg.norm(points[1] - points[2])))
pts_std = np.float32([[0, 0], [img_crop_width, 0],
[img_crop_width, img_crop_height],
[0, img_crop_height]])
M = cv2.getPerspectiveTransform(points, pts_std)
dst_img = cv2.warpPerspective(
img,
M, (img_crop_width, img_crop_height),
borderMode=cv2.BORDER_REPLICATE,
flags=cv2.INTER_CUBIC)
dst_img_height, dst_img_width = dst_img.shape[0:2]
if dst_img_height * 1.0 / dst_img_width >= 1.5:
dst_img = np.rot90(dst_img)
return dst_img
def ocr(self, image_path):
"""
完整的OCR流程
"""
# 读取图像
image = cv2.imread(image_path)
if image is None:
return []
# 1. 文本检测
dt_boxes, scores = self.detect_text(image)
# 可视化检测框
self.visualize_det_results(image_path, dt_boxes)
if dt_boxes is None or len(dt_boxes) == 0:
return []
# 2. 文本识别
ocr_results = []
text_list = []
confidence_list = []
for i, box in enumerate(dt_boxes):
# 裁剪文本区域
box_points = box.astype(np.float32)
crop_img = self.get_rotate_crop_image(image, box_points)
# 识别文本
text, conf = self.recognize_text(crop_img)
if conf > 0.4: # 置信度过滤
ocr_results.append({
'text': text,
'confidence': conf,
'box': box.tolist(),
'score': scores[i] if i < len(scores) else 0.0
})
text_list.append(text)
confidence_list.append(round(conf.item(), 2))
# return ocr_results
return [text_list, confidence_list]
def release(self):
"""释放资源"""
self.det_rknn.release()
self.rec_rknn.release()
def main():
parser = argparse.ArgumentParser(description='RK3588 PaddleOCR RKNN推理')
parser.add_argument('--det_model', type=str, required=True, help='文本检测RKNN模型路径')
parser.add_argument('--rec_model', type=str, required=True, help='文本识别RKNN模型路径')
parser.add_argument('--image', type=str, help='输入图像路径')
parser.add_argument('--video', type=str, help='输入视频路径')
parser.add_argument('--camera', type=int, help='摄像头设备ID')
parser.add_argument('--output', type=str, help='输出路径')
parser.add_argument('--show', action='store_true', help='显示结果')
args = parser.parse_args()
# 检查模型文件
if not os.path.exists(args.det_model):
print(f"检测模型文件不存在: {args.det_model}")
return
if not os.path.exists(args.rec_model):
print(f"识别模型文件不存在: {args.rec_model}")
return
# 初始化OCR
print("初始化RK3588 OCR...")
ocr = RK3588OCR(args.det_model, args.rec_model)
try:
if args.image:
# 图像模式
print(f"处理图像: {args.image}")
# 进行OCR
start_time = time.time()
text, confidence = ocr.ocr(args.image)
print("text", text)
print("confidence", confidence)
total_time = time.time() - start_time
# 打印结果
print(f"\n总耗时: {total_time:.3f}s")
print(f"识别结果:")
for i in range(len(text)):
print(f"{i+1}. 文本: '{text[i]}', 置信度: {confidence[i]:.3f}")
# 绘制结果
# annotated_image = ocr.draw_results(image, results)
# 保存或显示结果
# if args.output:
# cv2.imwrite(args.output, annotated_image)
# print(f"结果已保存到: {args.output}")
# if args.show:
# cv2.imshow('OCR结果', annotated_image)
# cv2.waitKey(0)
# cv2.destroyAllWindows()
elif args.video or args.camera is not None:
# 视频或摄像头模式
if args.video:
cap = cv2.VideoCapture(args.video)
print(f"处理视频: {args.video}")
else:
cap = cv2.VideoCapture(args.camera)
print(f"使用摄像头: {args.camera}")
if not cap.isOpened():
print("无法打开视频源")
return
while True:
ret, frame = cap.read()
if not ret:
break
# 进行OCR
results = ocr.ocr(frame)
# 绘制结果
annotated_frame = ocr.draw_results(frame, results)
# 显示结果
cv2.imshow('Real-time OCR', annotated_frame)
# 按'q'退出
if cv2.waitKey(1) & 0xFF == ord('q'):
break
cap.release()
cv2.destroyAllWindows()
else:
print("请指定输入源: --image, --video 或 --camera")
finally:
ocr.release()
if __name__ == "__main__":
# 如果直接运行脚本，提供示例用法
'''
启动命令示例
python 010使用PaddleOCR_rknn.py --det_model ./text_detection.rknn --rec_model ./text_recognition.rknn --image ./image_test/632e474452d560edd7004f745319ff00_frame_000730.jpg --output ./result.jpg
'''
if len(os.sys.argv) == 1:
print("RK3588 PaddleOCR RKNN推理程序")
print("\n使用示例:")
print("# 处理单张图像")
print("python rk3588_ocr.py \\")
print(" --det_model ./rknn_models/text_detection.rknn \\")
print(" --rec_model ./rknn_models/text_recognition.rknn \\")
print(" --image ./test.jpg \\")
print(" --output ./result.jpg \\")
print(" --show")
print()
print("# 实时摄像头OCR")
print("python rk3588_ocr.py \\")
print(" --det_model ./rknn_models/text_detection.rknn \\")
print(" --rec_model ./rknn_models/text_recognition.rknn \\")
print(" --camera 0")
print()
print("# 处理视频文件")
print("python rk3588_ocr.py \\")
print(" --det_model ./rknn_models/text_detection.rknn \\")
print(" --rec_model ./rknn_models/text_recognition.rknn \\")
print(" --video ./input_video.mp4")
else:
main()