deepellipse-circleid / code /color_aug_class.py
qinxunhui's picture
Add small artifacts (code, outputs, configs, LICENSE)
519ab98 verified
Raw History Blame Contribute Delete
11.8 kB
import cv2
import numpy as np
import os
import random
import matplotlib.pyplot as plt
import pickle
import json
from tqdm import tqdm
# ================= 辅助函数 (二值化与骨架化) =================
def binarize_image(image: np.ndarray, method: str = 'adaptive', k=1) -> np.ndarray:
"""
k=2,默认
k=0,含有更多边缘
"""
if len(image.shape) == 3:
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
else:
gray = image.copy()
if method == 'adaptive':
# 优化参数以保留更多细节
return cv2.adaptiveThreshold(
gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY_INV, 15, k
)
elif method == 'otsu':
_, binary = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
return binary
else:
_, binary = cv2.threshold(gray, 100, 255, cv2.THRESH_BINARY_INV)
return binary
def create_hard_mask(img):
"""生成硬二值Mask (0或255),用于确定变换区域"""
if len(img.shape) == 3:
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
else:
gray = img
blur = cv2.GaussianBlur(gray, (5, 5), 0)
ret, mask = cv2.threshold(blur, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
kernel = np.ones((3, 3), np.uint8)
mask = cv2.morphologyEx(mask, cv2.MORPH_OPEN, kernel, iterations=1)
mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernel, iterations=2)
if cv2.countNonZero(mask) / mask.size < 0.01:
_, mask = cv2.threshold(blur, 200, 255, cv2.THRESH_BINARY_INV)
mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernel, iterations=2)
# kernel = np.ones((3, 3), np.uint8)
# mask = cv2.erode(mask, kernel, iterations=2)
return mask
class ColorStatsDatabase:
"""
负责预计算、保存和加载参考图像库的统计信息。
针对白底黑字场景,仅计算'前景区域'的均值和方差。
"""
def __init__(self, cache_file='test_color_sta.json'):
# 缓存文件保存在参考文件夹的父目录或当前目录,避免污染图片文件夹
self.cache_file = cache_file
self.stats_data = {}
self.file_list = []
if os.path.exists(self.cache_file):
self.load()
else:
raise ValueError("输入正确的资源文件路径")
def compute_and_save(self, folder_path):
self.folder_path = folder_path
"""遍历文件夹,计算统计量并保存到磁盘"""
print(f"正在扫描并计算参考库统计量: {self.folder_path} ...")
valid_ext = ('.jpg', '.jpeg', '.png')
files = [f for f in os.listdir(self.folder_path)
if f.lower().endswith(valid_ext) and f.startswith('v2_')]
if not files:
raise ValueError(f"文件夹中未找到图像: {self.folder_path}")
self.stats_data = {}
for filename in tqdm(files, desc="计算统计量"):
filepath = os.path.join(self.folder_path, filename)
img = cv2.imread(filepath)
if img is None:
continue
mask = create_hard_mask(img)
mask_bool = mask > 0
if np.sum(mask_bool) < 10:
continue # 跳过无前景的图
means = []
stds = []
# 分别计算 B, G, R 通道的前景统计量
for i in range(3):
channel = img[:, :, i]
pixels = channel[mask_bool].astype(np.float64)
means.append(float(np.mean(pixels)))
stds.append(float(np.std(pixels)))
self.stats_data[filename] = {
'mean': means,
'std': stds
}
with open(self.cache_file, 'w') as f:
json.dump(self.stats_data, f, indent=2)
print(f"统计量已保存至: {self.cache_file}")
print(f"共处理有效图像: {len(self.stats_data)} 张")
self.file_list = list(self.stats_data.keys())
def load(self):
"""从磁盘加载统计量"""
if not os.path.exists(self.cache_file):
print(f"缓存文件不存在: {self.cache_file},将重新计算...")
self.compute_and_save()
else:
print(f"正在加载缓存统计量: {self.cache_file}")
with open(self.cache_file, 'r') as f:
self.stats_data = json.load(f)
self.file_list = list(self.stats_data.keys())
print(f"加载完成,共 {len(self.stats_data)} 张图像统计量。")
def get_random_stats(self):
"""随机返回一组统计量 (filename, mean, std)"""
if not self.file_list:
raise ValueError("数据库为空")
fname = random.choice(self.file_list)
data = self.stats_data[fname]
return fname, data['mean'], data['std']
class CircleColorAugmentor:
"""
颜色增强类:利用预计算的统计量对源图像进行硬变换(无边缘融合)。
"""
def __init__(self, ref_stats_db: ColorStatsDatabase, match_std=True):
self.ref_db = ref_stats_db
self.match_std = match_std
def augment(self, src_img, b_debug=False):
"""
执行一次增强:
1. 读取源图
2. 随机选择一组参考统计量
3. 生成源图硬Mask
4. 计算源图前景统计量
5. 应用线性变换并直接替换前景像素
"""
# src_img = cv2.imread(src_img_path)
# channels = cv2.split(src_img)
# kernel = np.ones((3, 3), np.uint8)
# new_ch = []
# for ch in channels:
# mod = cv2.erode(ch, kernel, iterations=2)
# new_ch.append(mod)
# src_img = cv2.merge(new_ch)
# 1. 获取随机参考统计量 (从内存缓存读取,极速)
ref_fname, ref_means, ref_stds = self.ref_db.get_random_stats()
# 2. 生成源图的硬 Mask (必须实时生成)
hard_mask = create_hard_mask(src_img)
mask_bool = hard_mask > 0
if np.sum(mask_bool) < 10:
# 没检测到前景,直接返回原图和空mask
if b_debug is True:
return src_img, src_img.copy(), np.zeros_like(hard_mask), "No_Foreground"
else:
return src_img
# 3. 计算源图当前前景的统计量
src_means = []
src_stds = []
for i in range(3):
channel = src_img[:, :, i]
pixels = channel[mask_bool].astype(np.float64)
src_means.append(np.mean(pixels))
src_stds.append(np.std(pixels))
# 4. 应用变换 (直接替换,无融合)
result_img = src_img.copy().astype(np.float64)
for i in range(3):
s_chan = src_img[:, :, i].astype(np.float64)
m_s, s_s = src_means[i], src_stds[i]
m_r, s_r = ref_means[i], ref_stds[i]
# 提取前景像素
src_vals = s_chan[mask_bool]
# 计算新值
if s_s < 1e-5:
new_vals = np.full_like(src_vals, m_r)
else:
if self.match_std:
# 匹配均值和方差
new_vals = (src_vals - m_s) * (s_r / s_s) + m_r
else:
# 仅匹配均值
new_vals = src_vals - m_s + m_r
# 截断
new_vals = np.clip(new_vals, 0, 255)
# 【关键】直接赋值给结果图像的 Mask 区域,背景保持不变
result_img[:, :, i][mask_bool] = new_vals
result_img = np.clip(result_img, 0, 255).astype(np.uint8)
# channels = cv2.split(result_img)
# new_ch = []
# for ch in channels:
# mod = cv2.dilate(ch, kernel, iterations=2)
# new_ch.append(mod)
# result_img = cv2.merge(new_ch)
# result_img = cv2.GaussianBlur(result_img, (5, 5), 0)
if b_debug is True:
return src_img, result_img, hard_mask, ref_fname
else:
return result_img
def visualize_results(src_img, tgt_img, hard_mask, ref_path, save_path):
"""生成对比图:源图 | 结果图 | 硬Mask | 参考信息"""
src_rgb = cv2.cvtColor(src_img, cv2.COLOR_BGR2RGB)
tgt_rgb = cv2.cvtColor(tgt_img, cv2.COLOR_BGR2RGB)
# 硬Mask 可视化
mask_vis = hard_mask.astype(np.uint8)
mask_rgb = cv2.cvtColor(mask_vis, cv2.COLOR_GRAY2RGB)
fig, axes = plt.subplots(1, 4, figsize=(20, 5))
axes[0].imshow(src_rgb)
axes[0].set_title("Source Image", fontsize=14, fontweight='bold')
axes[0].axis('off')
axes[1].imshow(tgt_rgb)
axes[1].set_title(f"Target (Hard Transform)", fontsize=14, fontweight='bold')
axes[1].axis('off')
axes[2].imshow(mask_rgb)
axes[2].set_title("Foreground Mask (Binary)", fontsize=14, fontweight='bold')
axes[2].axis('off')
# 第四列显示参考信息
ref_rgb = cv2.imread(ref_path)
axes[3].imshow(ref_rgb)
axes[3].set_title(f"Reference")
axes[3].axis('off')
plt.tight_layout()
plt.savefig(save_path, dpi=150, bbox_inches='tight')
plt.close(fig)
def debug():
# ================= 配置区域 =================
path_src_folder = "/path/to/circle_id/images/ours_images"
path_ref_folder = "/path/to/circle_id/images/images/"
path_out_folder = "/path/to/circle_id/outputs/color_aug_hard_class_new"
cache_file = os.path.join(os.path.dirname(__file__), "test_color_sta.json")
ref_db = ColorStatsDatabase(cache_file)
# test_color_sta.compute_and_save('/path/to/circle_id/images')
MATCH_STANDARD_DEVIATION = False # 是否匹配方差
NUM_SAMPLES = 100 # 生成测试图数量
# ===========================================
os.makedirs(path_out_folder, exist_ok=True)
# 2. 初始化增强器
augmentor = CircleColorAugmentor(ref_db, match_std=MATCH_STANDARD_DEVIATION)
# 3. 获取源图像列表
valid_ext = ('.jpg', '.jpeg', '.png')
src_files = [f for f in os.listdir(path_src_folder) if f.lower().endswith(valid_ext)]
if not src_files:
print("源文件夹为空")
return
print(f"\n开始生成 {NUM_SAMPLES} 张增强图像 (硬变换模式)...\n")
for idx in range(NUM_SAMPLES):
src_fname = random.choice(src_files)
src_path = os.path.join(path_src_folder, src_fname)
try:
src_img, tgt_img, hard_mask, ref_used = augmentor.augment(src_path, b_debug=True)
if ref_used == "No_Foreground":
print(f"[{idx + 1}/{NUM_SAMPLES}] 跳过 (未检测到前景): {src_fname}")
continue
# 保存结果
out_name = f"aug_{idx:03d}_{os.path.splitext(src_fname)[0]}_ref_{os.path.splitext(ref_used)[0]}.jpg"
save_path = os.path.join(path_out_folder, out_name)
ref_used = path_ref_folder + ref_used
visualize_results(src_img, tgt_img, hard_mask, ref_used, save_path)
print(f"[{idx + 1}/{NUM_SAMPLES}] 完成: {out_name}")
except Exception as e:
print(f"[{idx + 1}/{NUM_SAMPLES}] 失败: {e}")
import traceback
traceback.print_exc()
print("\n所有任务完成!结果已保存。")
def main():
cache_file = os.path.join(os.path.dirname(__file__), "test_color_sta.json")
ref_db = ColorStatsDatabase(cache_file)
augmentor = CircleColorAugmentor(ref_db, match_std=True)
src_path = "/path/to/circle_id/images/ours_images/W103_17_circle_010.png"
tgt_img = augmentor.augment(src_path, b_debug=False)
cv2.imwrite("/path/to/circle_id/outputs/tgt.png", tgt_img)
if __name__ == "__main__":
# main()
debug()