该代码段展示了一个Python脚本，用于处理高分辨率图像中的小目标检测。它首先使用Labelme工具进行标注，然后找到标注区域的最小包围框，计算中心并裁剪出640*640的训练样本。如果目标位于边界，脚本会调整边界线以适应。脚本还涉及图像的平移、缩放和坐标转换，以确保裁剪后的图像和对应的标注文件正确无误。
两千万像素的图，要检测的东西很小，可以先用labelme标注
测试图像
然后通过以找到标注区域的最小包围框，再算出包围框中心，从而裁切出640*640的训练样本图片，如果在检测对象在框的边界，则通过连接离边界最近的两点和之前的多边形形成标注。这种处理接近边界的情况，不是很理想，可以通过直接用边界的线替代超出边界的部分。
import numpy as np
import os
import cv2
import xml.etree.ElementTree as ET
import shutil
import random
import json
import base64
from PIL import Image
from copy import deepcopy
np.random.seed(1337) # 可重现随机数
import shutil
# ul左上角，lr右下角
def pingyi(pic_ul_point, pic_lr_point): # 左上角，右下角
if pic_ul_point[0] < 0:
pic_lr_point[0] += abs(0 - pic_ul_point[0])
pic_ul_point[0] = 0
if pic_ul_point[1] < 0:
pic_lr_point[1] += abs(0 - pic_ul_point[1])
pic_ul_point[1] = 0
# 图像最大值
if pic_lr_point[0] > 5472:
pic_ul_point[0] -= abs(5472 - pic_lr_point[0])
pic_lr_point[0] = 5472
if pic_lr_point[1] > 3648:
pic_ul_point[1] -= abs(3648 - pic_lr_point[1])
pic_lr_point[1] = 3648
return pic_ul_point, pic_lr_point
def pdInPic(shapes_list_i2, pic_ul_point, pic_lr_point):
points = shapes_list_i2["points"]
x_min, y_min, x_max, y_max = points[0][0], points[0][1], points[0][0], points[0][1]
for points_i in points:
if x_min > points_i[0]:
x_min = points_i[0]
if x_max < points_i[0]:
x_max = points_i[0]
if y_min > points_i[1]:
y_min = points_i[1]
if y_max < points_i[1]:
y_max = points_i[1]
xmid, ymid = int((x_max - x_min) / 2 + x_min), int((y_max - y_min) / 2 + y_min)
flag_shapeInArea = 0
if (pic_ul_point[0] < xmid < pic_lr_point[0]) and (pic_ul_point[1] < ymid < pic_lr_point[1]):
flag_shapeInArea = 1
return flag_shapeInArea
def crop640(path_input, path_output):
# 设置图片和标注文件列表
# imgs_list = os.listdir(path_input + "/imgs")
# jsons_list = os.listdir(path_input + "/jsons")
imgs_list = os.listdir(path_input)
jsons_list = os.listdir(path_input)
# imgs_outPath = save_path + "/imgs"
# jsons_outPath = save_path + "jsons"
for i in imgs_list:
if i.endswith((".png", ".jpg", ".bmp", ".jpeg")):
j = i[:-4] + ".json"
if j in jsons_list:
# 读取图片，计算分割后w,h
img0 = cv2.imread(path_input + "/" + i, 1)
w0, h0 = img0.shape[1], img0.shape[0]
w, h = 640, 640
# 读取原来json内容
f0 = open(path_input + '/' + j, "r", encoding="utf-8")
json_file = json.load(f0)
shapes_list = json_file["shapes"]
y_version = json_file["version"]
f0.close()
# 修改并保存
idex = 0
for shapes_list_i in shapes_list: # 单个缺陷
idex += 1
points = shapes_list_i["points"]
new_shapes_list = []
# 判断缺陷最大包围框是否大于640*640
flagIn_640 = 0
x_min, y_min, x_max, y_max = points[0][0], points[0][1], points[0][0], points[0][1]
for points_i in points:
if x_min > points_i[0]:
x_min = points_i[0]
if x_max < points_i[0]:
x_max = points_i[0]
if y_min > points_i[1]:
y_min = points_i[1]
if y_max < points_i[1]:
y_max = points_i[1]
if (y_max-y_min) < h and (x_max-x_min) < w:
flagIn_640 = 1
#
# cv2.rectangle(img0, (int(x_min), int(y_min)), (int(x_max), int(y_max)), (0, 255, 0), 1)
# cv2.imwrite('./linshi01.png', img0)
# # cv2.imshow("imge", img0)
# # cv2.waitKey()
# 缺陷小于640，裁剪周围640*640区域，只保存缺陷中心点在图片内的缺陷，其余忽略，保存新图片
# 标注同时改变坐标，保存新txt
if flagIn_640:
xmid, ymid = int((x_max-x_min)/2+x_min), int((y_max-y_min)/2+y_min)
pic_ul_point, pic_lr_point = [xmid-320, ymid-320], [xmid+320, ymid+320]
# 如果裁剪图片超过原图，平移
pic_ul_point, pic_lr_point = pingyi(pic_ul_point, pic_lr_point)
shapes_list2 = deepcopy(shapes_list)
for shapes_list_i2 in shapes_list2:
# 判断单个缺陷是否全部包含在图片内
flagInPic = pdInPic(shapes_list_i2, pic_ul_point, pic_lr_point)
if flagInPic:
new_shapes_list.append(shapes_list_i2)
# 保存裁剪图片
crop_img = img0[pic_ul_point[1]:pic_lr_point[1], pic_ul_point[0]:pic_lr_point[0]]
cv2.imwrite(path_output + '/' + j[:-4] + "_" + str(idex) + '.png', crop_img)
# 改变坐标，保存标注文件
new_shapes_list2 = []
for shapes_list_i3 in new_shapes_list:
points = shapes_list_i3["points"]
new_Points = []
for point in points:
new_point = [point[0]-pic_ul_point[0], point[1]-pic_ul_point[1]]
# 将在图片外的点去除
if (0 < new_point[0] < (pic_lr_point[0]-pic_ul_point[0])) and (0 < new_point[1] < (pic_lr_point[1] - pic_ul_point[1])):
new_Points.append(new_point)
shapes_list_i3["points"] = new_Points
new_shapes_list2.append(shapes_list_i3)
cur_json_dict = {
"version": y_version,
"flags": {},
}
newImgPath = path_output + '/' + j[:-4] + "_" + str(idex) + '.png'
cur_json_dict['shapes'] = new_shapes_list2
cur_json_dict['imagePath'] = newImgPath
cur_json_dict["imageData"] = str(base64.b64encode(open(newImgPath, "rb").read()))
cur_json_dict["imageData"] = cur_json_dict["imageData"][2:-1]
cur_json_dict["imageHeight"] = h
cur_json_dict["imageWidth"] = w
# if shapes_list_i
with open(newImgPath[:-4] + ".json", 'w') as f:
f.write(json.dumps(cur_json_dict))
f.close()
# 缺陷大于640，按最长边裁剪矩形框，如果矩形框大于3648则短边为3648，将裁剪图片缩放成640*640，保存新图片
# 标注同时改变坐标，按比例缩放，保存新txt
else:
pic_ul_point, pic_lr_point = [int(x_min-20), int(y_min-20)], [int(x_max+20), int(y_max+20)]
# 如果裁剪图片超过原图，平移
pic_ul_point, pic_lr_point = pingyi(pic_ul_point, pic_lr_point)
shapesList2 = deepcopy(shapes_list)
for shapes_list_i2 in shapesList2:
# 判断单个缺陷是否全部包含在图片内
flagInPic = pdInPic(shapes_list_i2, pic_ul_point, pic_lr_point)
if flagInPic:
new_shapes_list.append(shapes_list_i2)
# 缩放，保存裁剪图片
crop_img = img0[pic_ul_point[1]:pic_lr_point[1], pic_ul_point[0]:pic_lr_point[0]]
crop_img = cv2.resize(crop_img, (w, h), interpolation=cv2.INTER_AREA)
cv2.imwrite(path_output + '/' + j[:-4] + "_" + str(idex) + '.png', crop_img)
# 改变坐标，保存标注文件
sf_ratio_x, sf_ratio_y = (pic_lr_point[0]-pic_ul_point[0]) / w, (pic_lr_point[1]-pic_ul_point[1]) / h
new_shapes_list2 = []
for shapes_list_i3 in new_shapes_list:
points = shapes_list_i3["points"]
new_Points = []
for point in points:
new_point = [(point[0] - pic_ul_point[0]) / sf_ratio_x, (point[1] - pic_ul_point[1]) / sf_ratio_y]
# 将在图片外的点去除
if (0 < new_point[0] < (pic_lr_point[0] - pic_ul_point[0])) and (0 < new_point[1] < (pic_lr_point[1] - pic_ul_point[1])):
new_Points.append(new_point)
shapes_list_i3["points"] = new_Points
new_shapes_list2.append(shapes_list_i3)
cur_json_dict = {
"version": y_version,
"flags": {},
}
newImgPath = path_output + '/' + j[:-4] + "_" + str(idex) + '.png'
cur_json_dict['shapes'] = new_shapes_list2
cur_json_dict['imagePath'] = newImgPath
cur_json_dict["imageData"] = str(base64.b64encode(open(newImgPath, "rb").read()))
cur_json_dict["imageData"] = cur_json_dict["imageData"][2:-1]
cur_json_dict["imageHeight"] = h
cur_json_dict["imageWidth"] = w
# if shapes_list_i
with open(newImgPath[:-4] + ".json", 'w') as f:
f.write(json.dumps(cur_json_dict))
f.close()
if __name__ == '__main__':
path_input = "./test_crop/input"
path_output = "./test_crop/output"
crop640(path_input, path_output)
运行裁切完后
扫一扫
8285
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
