mask rcnn Keras+Tensor 训练自己的数据集

mask rcnn Keras+Tensor 训练自己的数据集

前言

最近在使用mask rcnn训练自己的数据,主要使用的是基于Mask RCNN开源项目github地址为
https://github.com/matterport/Mask_RCNN
本文主要参考
https://blog.csdn.net/qq_29462849/article/details/81037343
https://blog.csdn.net/l297969586/article/details/79140840
https://blog.csdn.net/u012746060/article/details/82143285
但是采用以上博客的方法的时候会遇到问题:

1. 重写的draw_mask功能,采用了两层for循环来遍历像素,这样生成mask的时候会非常耗时；
2. 重写的loadmask的功能,直接将# number of object=1,这个不太明白,这样反馈出来的mask的层数好像不对;
3. 重新写draw_mask的设置设置像素为1的时候,只用了三维；
4. 以上的问题,应该用示例代码中的## Load and display random samples功能来验证一下,由于mask实在load不出来；

重新写laodmask功能

参考了官方的例程Balloon.py和train_shapes两个例子程序

1. Balloon程序的标注格式是使用VIA标注,Json文件是将所有的图片保存到一个.json文件,文件的格式参考如下

// annotations
        # VGG Image Annotator (up to version 1.6) saves each image in the form:
        # { 'filename': '28503151_5b5b7ec140_b.jpg',
        #   'regions': {
        #       '0': {
        #           'region_attributes': {},
        #           'shape_attributes': {
        #               'all_points_x': [...],
        #               'all_points_y': [...],
        #               'name': 'polygon'}},
        #       ... more regions ...
        #   },
        #   'size': 100202
        # }

2. add_image可以发现这个是添加了polygon到imagelist里面
3. 其实labelme的json的数据也有polygon的信息,只不过存储的方式不一样,所以可以先将json的shape中代码如下

// annotations
        for json_file in imglist:
            with json_file.open() as f:
                json_result = json.load(f)
            if type(json_result['shapes']) is dict:
                polygons = [r['points'] for r in json_result['shapes'].values()]
                shapes=[r['label'] for r in json_result['shapes']]
            else:
                polygons = [r['points'] for r in json_result['shapes']]
                shapes=[r['label'] for r in json_result['shapes']]
                #shapes=

            # load_mask() needs the image size to convert polygons to masks.
            # Unfortunately, VIA doesn't include it in JSON, so we must read
            # the image. This is only managable since the dataset is tiny.
            # labelme include the height and weight

            image_path = os.path.join(dataset_dir, json_result['imagePath'])
            height=json_result['imageHeight']
            width = json_result['imageWidth']

            self.add_image(
                "shapes",
                image_id=json_result['imagePath'],  # use file name as a unique image id
                path=image_path,
                width=width, height=height,
                polygons=polygons,
                shapes=shapes)

4. 通过loadmask可以通过将mask按照类别添加到mask中代码如下

    def load_mask(self, image_id):
        """Generate instance masks for an image.
       Returns:
        masks: A bool array of shape [height, width, instance count] with
            one mask per instance.
        class_ids: a 1D array of class IDs of the instance masks.
        """
        # If not a balloon dataset image, delegate to parent class.
        #image_info = self.image_info[image_id]
        #if image_info["source"] != "balloon":
        #    return super(self.__class__, self).load_mask(image_id)
        info = self.image_info[image_id]
        shapes = info['shapes']
        count = len(shapes) # number of object

        # Convert polygons to a bitmap mask of shape
        # [height, width, instance_count]
        
        mask = np.zeros([info["height"], info["width"], count],
                        dtype=np.uint8)
        for i, p in enumerate(info['polygons']):        
            p_y=[]
            p_x=[]
            for point in p:
                p_y.append(point[1])
                p_x.append(point[0])
            rr, cc = skimage.draw.polygon(p_y, p_x)
            mask[rr, cc, i:i+1] = 1

        # Handle occlusions
        occlusion = np.logical_not(mask[:, :, -1]).astype(np.uint8)
        for i in range(count-2, -1, -1):
            mask[:, :, i] = mask[:, :, i] * occlusion
            occlusion = np.logical_and(occlusion, np.logical_not(mask[:, :, i]))

        # Map class names to class IDs.
        class_ids = np.array([self.class_names.index(s) for s in shapes])
        return mask.astype(np.bool), class_ids.astype(np.int32)

5. 用display_top_masks显示出来的效果
   

训练自己的数据

之后就可以愉快的训练自己的数据了