【二值标注maks图转coco的json标注，两个json文件合并，检查错误标注】

自定义数据集转成json格式，和原始数据集合并训练maskRCNN网络

首先是二值标注图转成coco类型 mask_to_coco。

这里我的自定义数据集有原始数据集中的两个类别Dirty和Scratch，
每张图像只包含一个类别

// An highlighted block
def mask_to_coco():
    # for categ in category:
    #     path = r"E:\Documents\datasets\VISION\round1\mask/%s" % categ
    for index, i in enumerate(tqdm(os.listdir(path))):
        if i.split('.')[-1] != 'png':
            continue
        mask_file = os.path.join(path, i)
        name = i.split('.')[0]
        file = os.path.join(path, '{}.png'.format(name))

        image = cv2.imread(file)
        if image.sum() / 255 < 10:
            continue
        image = Image(image, id=int(index + 1))  # 定义一个Image对象
        image.file_name = '{}.png'.format(name)  # 为上面的Image对象添加coco标签格式的'file_name'属性
        # image.path = file # 为Image对象添加coco标签格式的'path'属性

        mask = cv2.imread(mask_file, 0)
        t = cv2.imread(file)
        if t.shape[:-1] != mask.shape:
            h, w, _ = t.shape
            mask = cv2.resize(mask, (w, h), cv2.INTER_CUBIC)

        mask = Mask(mask)  # 定义一个Mask对象，并传入上面所定义的image对应的mask数组

        categ = i.split('_')[0]
        t = Category(categ)  # 这里是定义Category对象
        if categ == 'Dirty':
            t.id = 1  # 手动指定类别id号
            t.supercategory = 'Dirty'
        elif categ == 'Scratch':
            t.id = 3  # 同上
            t.supercategory = 'Scratch'
        # elif categ == 'removal':
        #     t.id = 3
        image.add(mask, t)  # 将mask信息和类别信息传给image
        dataset.add(image)  # 往dataset里添加图像以及gt信息

然后在主函数里面写入json文件：

dataset = Dataset('console')  # 先定义一个数据库对象，后续需要往里面添加具体的image和annotation
size = 800
path = r"E:\Documents\results\VISION\MDGAN\test\mask_"  # image对应的mask的文件路径
save_path = r"E:\Documents\results\VISION\MDGAN\test/"
N = 2227
annotation_id = 357
category = ['Dirty', 'Scratch', 'Collision', 'Gap']

if __name__ == '__main__':
    mask_to_coco()
    t = dataset.coco() # 将dataset转化为coco格式的，还可以转化为yolo等格式
    with open(save_path + '/mdgan_annotations.coco.json', 'w') as output_json_file: # 最后输出为json数据
         json.dump(t, output_json_file)
    compose_json()
    check_segmentation()

compose_json()用于将同一文件夹下的两个json文件组合起来：

def compose_json():
    path = r'E:\Documents\results\VISION\MDGAN\test/'
    filelist = os.listdir(path)
    real_json = r"E:\Documents\results\VISION\MDGAN\test\real_annotations.coco.json"
    main = open(real_json)
    main = json.load(main)
    num_images = len(main['images'])
    num_annos = len(main['annotations'])
    for entry in filelist:
        if entry.split('.')[-1] != 'json' or entry.find('real_annotations') != -1:
            continue
        print('Loading', entry)
        file = open(path + entry)
        file = json.load(file)

        for i in file['images']:
            main['images'].append(i)  # image id here is wrong
        for i in file['annotations']:
            for son_item in i['segmentation']:
                l = len(son_item)
                if l < 8:
                    i['segmentation'].remove(son_item)
                    print(son_item, 'is removed')
            main['annotations'].append(i)

        for i in range(len(main['images'])):
            main['images'][i]['id'] = i + 1  # resort
        for i in range(len(main['annotations'])):
            main['annotations'][i]['id'] = i + 1

        sameID = []
        for i in range(0, len(main['annotations'])):
            if (main['annotations'][i]['image_id'] != main['annotations'][i - 1]['image_id']):
                sameID.append(main['annotations'][i]['id'])

        for item in main['annotations'][num_annos:]:
            item['image_id'] += num_images

        # newList = []
        # c = 1
        # for i in range(len(sameID) - 1):
        #     newList.extend([c] * (sameID[i + 1] - sameID[i]))
        #     c = c + 1
        # # newList.extend([75, 75, 75, 75])
        # for i in range(len(newList)):
        #     main['annotations'][i]['image_id'] = newList[i]
    with open(path + '/_annotations.coco.json', 'w') as outfile:
        json.dump(main, outfile)

check_segmentation用于检测转换的json文件在segmentation时候有没有长度小于8的标注，如果有的话，训练时候会报错

ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (6,) + inhomogeneous part.
TypeError: Argument ‘bb’ has incorrect type (expected numpy.ndarray, got list)
Exception: input type is not supported.

def check_segmentation():
    savepath = r'E:\Documents\results\VISION\MDGAN\test/'
    filepath = savepath + '/1_annotations.coco.json'
    file = open(filepath)
    file = json.load(file)
    for item in file['annotations']:
        for son_item in item['segmentation']:
            l = len(son_item)
            if l < 8:
                item['segmentation'].remove(son_item)
                print(son_item, 'in', file['images'][item['image_id'] - 1]['file_name'], 'is removed')
        if len(item['segmentation']) < 1:
            file['annotations'].remove(item)
            print(item, 'in', file['images'][item['image_id']]['file_name'], 'is removed')
    with open(savepath + '/_annotations.coco.json', 'w') as outfile:
        json.dump(file, outfile)

我一共遇到了这三种错误，有时候程序能跑，但到epoch=100或者200左右时候就会报错，主要还是标注的格式问题，用check_segmentation检查清除之后就可以了。