半自动标注（使用自己的分割或检测模型推理完得到的矩阵再生成json文件）

本文主要是介绍半自动标注（使用自己的分割或检测模型推理完得到的矩阵再生成json文件），希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！

前言
都知道标注很麻烦、很累，不然先训练一批，然后推理得到它的掩码图，先生成自动标注，再人工手动修改也许会快很多

半自动标注代码

这是我自己写的，是labelme的格式，大家想要修改成自己的json格式可以修改json_dict_init与dict_init函数

默认大家已经得到自己图片经过模型推理之后的掩码图了，掩码图上面生成的像素应该是0,1,2,3. 分别对应自己训练的时候的标签

import cv2
import sys
import base64
import cv2
import json
import time
import os
from tqdm import tqdm
from PIL import Image
import numpy as npdef dict_init(label_name):temp_dict = {"label": label_name,"line_color": None,"fill_color": None,"points":[],"shape_type": "polygon","flags": {}}return  temp_dictdef json_dict_init(num,height,width,img):temp_dict = {"version": "3.16.2","flags": {},"shapes": [],"lineColor": [0,255,0,128],"fillColor": [255,0,0,128],"imagePath": "..\\img\\{}.jpg".format(num),#原图像数据通过b64编码生成的字符串数据，可以再次解码成图片"imageData":img,"imageHeight": height,"imageWidth": width}return  temp_dictdef image_to_base64(image_path):# 读取二进制图片，获得原始字节码with open(image_path, 'rb') as jpg_file:byte_content = jpg_file.read()# 把原始字节码编码成base64字节码base64_bytes = base64.b64encode(byte_content)# 把base64字节码解码成utf-8格式的字符串base64_string = base64_bytes.decode('utf-8')return base64_string# 定义一个函数，用于判断两个点是否相邻
def is_adjacent(pt1, pt2):x1, y1 = pt1x2, y2 = pt2return abs(x1 - x2) <= 8 and abs(y1 - y2) <= 8def main():img_path = "D:/data_val/new/temp/img"label_path = "D:/data_val/new/temp/label"json_save_path = 'D:/data_val/new/temp/json'os.makedirs(json_save_path,exist_ok=True)# 放入你的标签名 例如 你的掩码图 像素是1 1对应的就是phone 像素是2 2就是linelabel_name_list = ['phone','line']img_list = os.listdir(img_path)pbar = tqdm(total=len(img_list))for filename in img_list:name = filename.split(".")[0]# print("*"*10)# print(f'{filename}')#label图label = Image.open(f'{label_path}/{name}.png')label = np.array(label)#输出自己的label有多少种像素# unique_values = np.unique(label)# print('')# print(unique_values)# print('')height  = label.shape[0]width   = label.shape[1]base64_string = image_to_base64(f'{img_path}/{filename}')json_dict = json_dict_init(name,height,width,base64_string)contours, hierarchy = cv2.findContours(label, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)background_flag =Falseif (len(contours) == 0):print("如果未识别到 则变成背景")shapes_dict = dict_init("_background_")left_top, left_botton = [10, 10], [10, height - 10]right_botton, right_top = [width - 10, height - 10], [width - 10, 10]shapes_dict['points'] = [left_top, left_botton, right_botton, right_top]json_dict["shapes"].append(shapes_dict)else:for i,clasee_name in enumerate(label_name_list):temp = label.copy()temp[temp!=(i+1)] = 0contours, hierarchy = cv2.findContours(temp, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)if (len(contours) == 0):continueelse:count = 0for contour in contours:shapes_dict = dict_init(clasee_name)count = count + 1#10 20#语义分割approxCourve = cv2.approxPolyDP(contour, 3, True)for Courve in approxCourve:point = Courve[0]x, y = pointx = int(x)y = int(y)temp_list = []temp_list.append(x)temp_list.append(y)shapes_dict["points"].append(temp_list)#目标检测 这里只写到生成检测框的四个点 根据自己需求填进去# rect = cv2.minAreaRect(contour)# box = cv2.boxPoints(rect)# # print(box)# # 轮廓必须是整数, 不能是小数, 所以转化为整数# box = np.round(box).astype('int64')# left_point_x = np.min(box[:, 0])# right_point_x = np.max(box[:, 0])# top_point_y = np.min(box[:, 1])# bottom_point_y = np.max(box[:, 1])json_dict["shapes"].append(shapes_dict)with open(f'{json_save_path}/{name}.json', "w", encoding='utf-8') as f:f.write(json.dumps(json_dict, ensure_ascii=False))pbar.update(1)if __name__ == "__main__":main()