deepellipse-circleid / code /augmentation.py
qinxunhui's picture
Add small artifacts (code, outputs, configs, LICENSE)
519ab98 verified
Raw History Blame Contribute Delete
13.9 kB
import os
import cv2
import numpy as np
from pathlib import Path
import random
from circle_feature import exact_features
from color_aug_class import ColorStatsDatabase, CircleColorAugmentor
from image_precess import estimate_stroke_width
import traceback
DEF_PROB_DICT = {"crop": 0.3,
"rotation": 0.5,
"shear": 0.3,
"affine": 0.5,
"stroke": 0.3,
"color": 0.5,
"blur": 0.3
}
# DEF_PROB_DICT = {"crop": 1,
# "rotation": 1,
# "shear": 1,
# "affine": 1,
# "stroke": 1,
# "color": 1,
# "blur": 1
# }
# ================= 核心增强类 (带调试记录) =================
class CircleAugmentation:
def __init__(
self,
color_cache_file=os.path.join(os.path.dirname(__file__), "test_color_sta.json"),
prob_dict=DEF_PROB_DICT,
):
# 记录每一步的图像
self.steps = {}
ref_db = ColorStatsDatabase(color_cache_file)
self.color_augmentor = CircleColorAugmentor(ref_db, match_std=True)
self.prob_dict = prob_dict
def apply_crop(self, image):
"""步骤 1: 随机单方向边缘裁剪 (20% 概率)"""
h, w = image.shape[:2]
step_name = "1_Crop"
if random.random() > self.prob_dict["crop"]:
# 不发生裁剪,直接返回原图副本
self.steps[step_name] = image.copy()
self.steps[f"{step_name}_info"] = "No Crop"
return image
# 决定裁剪方向:0:Top, 1:Bottom, 2:Left, 3:Right
direction = random.randint(0, 3)
max_pixels = int(min(h, w) * 0.2)
crop_amount = random.randint(1, max_pixels)
y1, y2, x1, x2 = 0, h, 0, w
if direction == 0: # Top
y1 = crop_amount
info = f"Top-{crop_amount}px"
elif direction == 1: # Bottom
y2 = h - crop_amount
info = f"Bottom-{crop_amount}px"
elif direction == 2: # Left
x1 = crop_amount
info = f"Left-{crop_amount}px"
else: # Right
x2 = w - crop_amount
info = f"Right-{crop_amount}px"
cropped = image[y1:y2, x1:x2].copy()
self.steps[step_name] = cropped
self.steps[f"{step_name}_info"] = info
return cropped
def apply_rotation(self, image, degrees=20):
step_name = "rotation"
if random.random() > self.prob_dict["rotation"]:
# 不发生裁剪,直接返回原图副本
self.steps[step_name] = image.copy()
self.steps[f"{step_name}_info"] = "No " + step_name
return image
"""步骤 2: 随机旋转"""
h, w = image.shape[:2]
center = (w // 2, h // 2)
angle = np.random.uniform(-degrees, degrees)
M = cv2.getRotationMatrix2D(center, angle, 1.0)
rotated = cv2.warpAffine(image, M, (w, h), borderMode=cv2.BORDER_CONSTANT, borderValue=(255, 255, 255), flags=cv2.INTER_LINEAR)
self.steps["2_Rotate"] = rotated
self.steps["2_Rotate_info"] = f"rotate:{angle:.1f}"
return rotated
def apply_shear(self, image, degrees=10):
"""步骤 3: 随机剪切"""
step_name = "shear"
if random.random() > self.prob_dict["shear"]:
# 不发生裁剪,直接返回原图副本
self.steps[step_name] = image.copy()
self.steps[f"{step_name}_info"] = "No " + step_name
return image
h, w = image.shape[:2]
shear_x = np.random.uniform(-np.tan(np.radians(degrees)), np.tan(np.radians(degrees)))
shear_y = np.random.uniform(-np.tan(np.radians(degrees)), np.tan(np.radians(degrees)))
M = np.float32([[1, shear_x, 0], [shear_y, 1, 0]])
sheared = cv2.warpAffine(
image,
M,
(w, h),
borderMode=cv2.BORDER_CONSTANT, borderValue=(255, 255, 255),
flags=cv2.INTER_LINEAR
)
self.steps["3_Shear"] = sheared
self.steps["3_Shear_info"] = f"sheer-x:{shear_x:.2f}, y:{shear_y:.2f}"
return sheared
def apply_affine(self, image, translate=(0.05, 0.05), scale=(0.95, 1.05)):
"""步骤 4: 随机平移 + 缩放"""
step_name = "affine"
if random.random() > self.prob_dict["affine"]:
# 不发生裁剪,直接返回原图副本
self.steps[step_name] = image.copy()
self.steps[f"{step_name}_info"] = "No " + step_name
return image
h, w = image.shape[:2]
tx = np.random.uniform(-translate[0], translate[0]) * w
ty = np.random.uniform(-translate[1], translate[1]) * h
sx = np.random.uniform(scale[0], scale[1])
sy = np.random.uniform(scale[0], scale[1])
M = np.float32([[sx, 0, tx], [0, sy, ty]])
affine = cv2.warpAffine(image, M, (w, h), borderMode=cv2.BORDER_CONSTANT, borderValue=(255, 255, 255))
self.steps["4_Affine"] = affine
self.steps["4_Affine_info"] = f"affine aug"
return affine
def apply_stroke_aug(self, image):
"""步骤 5: 外观变换 (笔迹粗细)"""
if random.random() > self.prob_dict["stroke"]:
self.steps["5_Stroke"] = image.copy()
self.steps["5_Stroke_info"] = "No stroke aug"
return image
# 模拟动态逻辑 (这里简化为随机变粗或变细,实际可接入 width 估算)
kernel = np.ones((3, 3), np.uint8)
width = estimate_stroke_width(image)
if width > 6.5:
op_type = -1
# iter = random.randint(1, min(2, int((width - 3)/2)))
else:
op_type = 1
# iter = random.randint(1, min(2, int((7.5 - width)/2)))
iter = 1
if len(image.shape) == 3:
channels = cv2.split(image)
new_ch = []
for ch in channels:
if op_type == -1:
mod = cv2.dilate(ch, kernel, iterations=iter) #白色膨胀,笔画变细
else:
mod = cv2.erode(ch, kernel, iterations=iter) #白色腐蚀,笔画变粗
new_ch.append(mod)
result = cv2.merge(new_ch)
self.steps["5_Stroke"] = result
if op_type == -1:
self.steps["5_Stroke_info"] = "small:" + f"w:{width:.1f}"
else:
self.steps["5_Stroke_info"] = "large:" + f"w:{width:.1f}"
return result
def apply_color(self, image, use_prob=0.5):
"""步骤 5: 外观变换 (笔迹粗细)"""
if random.random() > use_prob:
self.steps["6_colar"] = image.copy()
self.steps["6_colar_info"] = "No color aug"
return image
image = self.color_augmentor.augment(image, b_debug=False)
self.steps["6_colar"] = image
self.steps["6_colar_info"] = "color aug"
return image
def apply_blur(self, image):
"""步骤 X: 外观变换 (图像模糊)"""
# 1. 概率控制:只有 use_prob (默认 30%) 的概率发生模糊
if random.random() > self.prob_dict["blur"]:
self.steps["7_Blur"] = image.copy()
self.steps["7_Blur_info"] = "No blur"
return image
ksize = random.choice([3, 5, 7])
blurred_image = cv2.GaussianBlur(image, (ksize, ksize), 0)
# 4. 记录调试信息
self.steps["7_Blur"] = blurred_image
self.steps["7_Blur_info"] = f"Gauss K={ksize}"
return blurred_image
def aug_image(self, image):
"""执行全流程并记录每一步"""
# 0. Original
self.steps["0_Original"] = image.copy()
self.steps["0_Original_info"] = "Input"
# 1. Crop
img = self.apply_crop(image)
# 2. Rotate
img = self.apply_rotation(img)
# 3. Shear
img = self.apply_shear(img)
# 4. Affine
img = self.apply_affine(img)
# 5. Stroke Aug
img = self.apply_stroke_aug(img)
# 6. color
img = self.apply_color(img)
# 7. blur
img = self.apply_blur(img)
return img
def aug_image_offical(self, img):
"""执行全流程并记录每一步"""
# 0. Original
self.steps["0_Original"] = img.copy()
self.steps["0_Original_info"] = "Input"
# 1. Crop
# img = self.apply_crop(image)
# 2. Rotate
img = self.apply_rotation(img)
# 3. Shear
# img = self.apply_shear(img)
# 4. Affine
img = self.apply_affine(img)
# 5. Stroke Aug
# img = self.apply_stroke_aug(img)
# # 6. color
# img = self.apply_color(img)
# # 7. blur
# img = self.apply_blur(img)
return img
def process_full_pipeline(self, image):
"""执行全流程并记录每一步"""
# 0. Original
self.steps["0_Original"] = image.copy()
self.steps["0_Original_info"] = "Input"
# 1. Crop
img = self.apply_crop(image)
# 2. Rotate
img = self.apply_rotation(img)
# 3. Shear
img = self.apply_shear(img)
# 4. Affine
img = self.apply_affine(img)
# 5. Stroke Aug
img = self.apply_stroke_aug(img)
# 6. color
img = self.apply_color(img)
# 7. blur
img = self.apply_blur(img)
return self.steps
# ================= 可视化工具 =================
def create_debug_image(steps, target_h=120):
"""将 steps 字典中的图像拼接成一行"""
row_images = []
# 定义显示顺序
order = [
"0_Original", "1_Crop", "2_Rotate", "3_Shear",
"4_Affine", "5_Stroke", "6_colar", "7_Blur"
]
for key in order:
if key not in steps:
continue
img = steps[key]
info = steps.get(f"{key}_info", "")
# 统一高度
h, w = img.shape[:2]
if h != target_h:
scale = target_h / h
new_w = int(w * scale)
img_resized = cv2.resize(img, (new_w, target_h))
else:
img_resized = img.copy()
# 添加底部文字标签
cv2.putText(img_resized, info, (5, target_h - 5),
cv2.FONT_HERSHEY_SIMPLEX, 0.4, (0, 255, 0), 1)
# 添加顶部标题 (简单用颜色条区分,或者只在第一张写)
if key == "0_Original":
cv2.putText(img_resized, "Original", (5, 15),
cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 2)
row_images.append(img_resized)
if not row_images:
return None
return np.hstack(row_images)
# ================= 主执行逻辑 =================
def debug_main():
INPUT_FOLDER = "/path/to/circle_id/images/images/"
OUTPUT_FOLDER = "/path/to/circle_id/outputs/augment2"
FILE_PATTERN = "*.png" # 或者 "*.png" 处理所有
NUM_SAMPLES = 100 # 调试几张图片
os.makedirs(OUTPUT_FOLDER, exist_ok=True)
print(f"✅ 输出目录:{OUTPUT_FOLDER}")
img_paths = list(Path(INPUT_FOLDER).glob(FILE_PATTERN))[:NUM_SAMPLES]
print(f"🔍 找到 {len(img_paths)} 张图片,开始生成调试图...")
aug = CircleAugmentation()
for idx, path in enumerate(img_paths):
# path = "/path/to/circle_id/images/images/10542.png"
# path = "/path/to/circle_id/images/images/14406.png"
# path = "/path/to/circle_id/images/images/21070.png"
# path = "/path/to/circle_id/images/images/30164.png"
#
# path = "/path/to/circle_id/images/images/38334.png"
# path = "/path/to/circle_id/images/images/39022.png"
path = "/path/to/circle_id/images/images/39890.png"
path_name = path.split("/")[-1]
# path_name = path.name
image = cv2.imread(str(path))
if image is None:
continue
all_rows = []
# 初始化增强器 (每次循环重新初始化以重置随机状态记录)
# 执行流程
steps = aug.process_full_pipeline(image)
# 生成拼接图
row_imgs = create_debug_image(steps, target_h=150)
if row_imgs is not None:
# 添加左侧文件名标签
h, w, _ = row_imgs.shape
label_img = np.ones((h, 150, 3), dtype=np.uint8) * 255
cv2.putText(label_img, path_name, (10, h // 2),
cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0), 1)
final_row = np.hstack([label_img, row_imgs])
all_rows.append(final_row)
# 垂直拼接所有行
# 确保宽度一致 (取最小宽度)
min_w = min([r.shape[1] for r in all_rows])
padded_rows = []
for r in all_rows:
if r.shape[1] > min_w:
r = r[:, :min_w, :]
padded_rows.append(r)
final_grid = np.vstack(padded_rows)
image_name = str(path).split("/")[-1]
output_path = os.path.join(OUTPUT_FOLDER, f"{idx}_{image_name}")
try:
feature = exact_features(steps["7_Blur"])
except Exception as e:
print(steps["7_Blur"].shape)
cv2.imwrite(output_path, final_grid)
print(f"\n🎉 错误调试图已保存:{output_path}")
print(str(path))
print(traceback.format_exc())
if __name__ == "__main__":
debug_main()
# src_path = "/path/to/circle_id/images/images/00151.png"
# img = cv2.imread(src_path)
# aug = CircleAugmentation()
# img_aug = aug.aug_image(img)
# cv2.imwrite("/path/to/circle_id/outputs/polar_order/temp/00151.png", img_aug)
# feature = exact_features(img_aug)