实例分割模型的数据集格式转换

在对生物医学图像进行处理时，不像自然图像有以前固定的数据集格式，比如coco的json格式，voc格式等等，而生物图像样式有点多且不标准，图片格式，npz格式等等。网上转换教程很多，但是其中不乏转换错误的代码，特此来记录，以下代码保证可以正常运行

1. coco格式

1.1 Yolo格式转COCO格式

参考博客：https://blog.csdn.net/ericdiii/article/details/137775872
下方代码来自上方博客（如有侵权立即删除），结合自己数据集需要修改的地方：
（1）图片后缀：默认是tif，根据自己是png，进行替换
（2）图片，标签，存储ann的路径

import json
import glob
import os
import cv2
import matplotlib.pyplot as plt
from matplotlib.patches import Polygon
from PIL import Image, ImageDraw, ImageFont
import numpy as npdef calculate_polygon_area(polygon):x = polygon[:, 0]y = polygon[:, 1]return 0.5 * np.abs(np.dot(x, np.roll(y, 1)) - np.dot(y, np.roll(x, 1)))def calculate_bounding_box(polygon):x_min = np.min(polygon[:, 0])y_min = np.min(polygon[:, 1])x_max = np.max(polygon[:, 0])y_max = np.max(polygon[:, 1])width = x_max - x_minheight = y_max - y_minreturn [x_min, y_min, width, height]def text_to_json_segmentation(in_labels, in_images, out_json):"""Convert instance segmentation dataset from text files generated by the function 'json_to_text_segmentation'(for YOLO) to a JSON file (for MMdet). This can be applied for Level 0/1/2 (must modify the last code):param in_labels: input folder containing the label text files:param in_images: input folder containing the image files (just for getting the image size):param out_json: output JSON file"""# Initialize the output JSON filedata = dict()data['annotations'] = []data['images'] = []# Initial the number of annotationsnum_annotations = 1  # index starts from 1# Process the text filestxt_files = glob.glob(in_labels + '/*.txt')for k in range(len(txt_files)):# Read the image to get image width and heightimg = Image.open(in_images + '/' + os.path.basename(txt_files[k]).replace('txt', 'tif'))image_width, image_height = img.size# Creates annotation items of the image and append them to the listwith open(txt_files[k]) as f:for line in f:# Get annotation information of each line in the text fileline = [float(x) for x in line.strip().split()]class_id = int(line[0]) + 1  # index starts from 1coordinates = line[1:]polygon = np.array(coordinates).reshape(-1, 2)polygon[:, 0] = polygon[:, 0] * image_widthpolygon[:, 1] = polygon[:, 1] * image_heightarea = calculate_polygon_area(polygon)bbox = calculate_bounding_box(polygon)# Create a new annotation itemann_item = dict()ann_item['segmentation'] = [polygon.flatten().tolist()]ann_item['area'] = areaann_item['iscrowd'] = 0ann_item['image_id'] = k + 1  # index starts from 1ann_item['bbox'] = bboxann_item['category_id'] = class_idann_item['id'] = num_annotationsdata['annotations'].append(ann_item)num_annotations += 1# Create a new image item and append it to the listimg_item = dict()img_item['id'] = k + 1  # index starts from 1img_item['file_name'] = os.path.basename(txt_files[k]).replace('txt', 'tif')img_item['height'] = image_heightimg_item['width'] = image_widthdata['images'].append(img_item)print(os.path.basename(txt_files[k]) + ' done')data['categories'] = [{'supercategory': 'point', 'id': 1, 'name': 'point'}]# Write the dictionary to a JSON fileprint('Writing the data to a JSON file')with open(out_json, 'w') as f:# json.dump(data, f, cls=NpEncoder)# f.write(json.dumps(data, cls=NpEncoder, indent=4))f.write(json.dumps(data, default=int, indent=4))if __name__ == '__main__':# Convert the segmentation text files to JSONtext_to_json_segmentation(in_labels='/home/liuhuan/pointseg/datasetsNormal/receptor/labels/valid',in_images='/home/liuhuan/pointseg/datasetsNormal/receptor/images/valid',out_json='/home/liuhuan/pointseg/datasetsNormal/receptor/annotations/instances_val2017.json')

如何对json标注文件进行检查：

生成的json文件不一定是对的，例如轮廓小于4个像素这种，需要检查，如果有错误，则会报错并打印错误图片的序号，结合自己数据集需要修改的是：
（1）json文件路径

import json
import numpy as np
from tqdm import tqdmdef frPyObjects(i, pyobj):# encode rle from a list of python objectsif type(pyobj) == np.ndarray:print("{}, {}, {}".format(i, type(pyobj), len(pyobj[0])))elif type(pyobj) == list and len(pyobj[0]) == 4:print("{}, {}, {}".format(i, type(pyobj), len(pyobj[0])))elif type(pyobj) == list and len(pyobj[0]) > 4:print("{}, {}, {}".format(i, type(pyobj), len(pyobj[0])))elif type(pyobj) == list and type(pyobj) == dict and 'counts' in pyobj[0] and 'size' in pyobj[0]:print("{}, {}, {}".format(i, type(pyobj), len(pyobj[0])))# encode rle from single python objectelif type(pyobj) == list and len(pyobj[0]) == 4:print("{}, {}, {}".format(i, type(pyobj), len(pyobj[0])))elif type(pyobj) == list and len(pyobj[0]) > 4:print("{}, {}, {}".format(i, type(pyobj), len(pyobj[0])))elif type(pyobj) == dict and 'counts' in pyobj and 'size' in pyobj:print("{}, {}, {}".format(i, type(pyobj), len(pyobj[0])))else:print("{}, {}, {},  ERROR".format(i, type(pyobj), len(pyobj[0])))raise Exception('input type is not supported.')def check(pathJson):jsonfile = open(pathJson, "r")jsonObj = json.load(jsonfile)jsonfile.close()for i, instance in tqdm(enumerate(jsonObj["annotations"])):frPyObjects(i, instance["segmentation"])if __name__ == "__main__":pathTrainJson = "/home/liuhuan/pointseg/datasetsNormal/receptor/annotations/instances_train2017.json"pathValJson = "/home/liuhuan/pointseg/datasetsNormal/receptor/annotations/instances_val2017.json"pathTestJson = "/home/liuhuan/pointseg/datasetsNormal/receptor/annotations/instances_test2017.json"check(pathTrainJson)check(pathValJson)check(pathTestJson)# 之后，可以根据上述print()打印出来的不符合条件对象的id，来修改对应的segmentation列表（手动修改即可，不符合条件的数量一般很少）
#json_object["annotations"][1510]["segmentation"] = [[230.83333333333331, 773.8888888888889, 231.83333333333331, 773.8888888888889, 237.22222222222223, 770.5555555555555]]# 将修改后的json文件重新写回到coco_instancestrain.json/coco_instancesval.json中即可
# val_json = open(JSON_LOC, "w")
# json.dump(json_object, val_json)
# val_json.close()

如何对json标注文件结合图片显示，进一步检查正确性

会根据图像和标注对某张图片显示，结合自己数据集需要修改的地方：
（1）图像文件所在文件夹
（2）json标注文件路径
（3）要查看图像文件的名称

import os.path
import numpy as np
import cv2
import json
from tqdm import tqdmdef get_bbox_segs(pathJson, img_name):jsonfile = open(pathJson, "r")jsonObj = json.load(jsonfile)jsonfile.close()# find img_id base img_namejsonImage = jsonObj["images"]i = 0while (jsonImage[i]["file_name"] != img_name):i = i + 1img_id = jsonImage[i]["id"]# find ann base img_idjsonAnn = jsonObj["annotations"]segs = []bboxes = []for i, instance in tqdm(enumerate(jsonAnn)):if instance["image_id"] == img_id:bboxes.append(instance["bbox"])segs.append(instance["segmentation"])return bboxes, segsif __name__ == "__main__":pathImg = "/home/liuhuan/pointseg/datasetsNormal/receptor/images/train"pathJson = "/home/liuhuan/pointseg/datasetsNormal/receptor/annotations/instances_train2017.json"img_name = "0.tif"bboxes, segs = get_bbox_segs(pathJson, img_name)img = cv2.imread(os.path.join(pathImg, img_name))# show bboxes'''for bbox in bboxes:x, y, w, h = bboxcv2.rectangle(img, (x, y), (x + w, y + h), (0, 255, 0), 2)'''# show masksmasks = []for seg in segs:mask = np.zeros((img.shape[0], img.shape[1]), dtype=np.uint8)polygon = np.array(seg).reshape((-1, 2)).astype(np.int32)cv2.fillPoly(mask, [polygon], 1)masks.append(mask)for mask in masks:mask_colored = cv2.cvtColor(mask*255, cv2.COLOR_GRAY2BGR)img = cv2.addWeighted(img, 1, mask_colored, 0.5, 0)cv2.imshow('Image', img)cv2.waitKey(0)cv2.destroyAllWindows()