图像预处理-大图切割-python实现

简介

深度学习中，数据集的预处理每每是很基础的一步，不少场景都须要将一张大图进行切割。本篇提供一种重叠矩形框的生成方法，数据集中的图像尺寸能够不一样，根据生成的重叠矩形框能够crop出相应的图像区域。主要难点在于函数不假设图像的尺寸大小。

实现

如下是重叠矩形框的生成函数，是根据右下角的坐标来肯定左上角的坐标，若是右下角的点超过了图像边缘，则让矩形的右下角等于边缘值。循环会让右下角的坐标往右和往下多走一个stride，这样能够将边缘部分的图像也包含进来。

#encoding=utf-8

def get_fixed_windows(image_size, wind_size, overlap_size):
    '''
    This function can generate overlapped windows given various image size
    params:
        image_size (w, h): the image width and height
        wind_size (w, h): the window width and height
        overlap (overlap_w, overlap_h): the overlap size contains x-axis and y-axis

    return:
        rects [(xmin, ymin, xmax, ymax)]: the windows in a list of rectangles
    '''
    rects = set()

    assert overlap_size[0] < wind_size[0]
    assert overlap_size[1] < wind_size[1]

    im_w = wind_size[0] if image_size[0] < wind_size[0] else image_size[0]
    im_h = wind_size[1] if image_size[1] < wind_size[1] else image_size[1]

    stride_w = wind_size[0] - overlap_size[0]
    stride_h = wind_size[1] - overlap_size[1]

    for j in range(wind_size[1]-1, im_h + stride_h, stride_h):
        for i in range(wind_size[0]-1, im_w + stride_w, stride_w):
            right, down = i+1, j+1
            right = right if right < im_w else im_w
            down  =  down if down < im_h  else im_h

            left = right - wind_size[0]
            up   = down  - wind_size[1]

            rects.add((left, up, right, down))      

    return list(rects)


if __name__ == "__main__":
    image_size = (1780, 532)
    wind_size = (800, 600)
    overlap_size = (300, 200)

    rets = get_fixed_windows(image_size, wind_size, overlap_size)

    for rect in rets:
        print(rect)
'''
# output
(0, 0, 800, 600)
(500, 0, 1300, 600)
(980, 0, 1780, 600)
'''

效果

总结

实在不知道写什么了，把以前项目里的一个图像预处理代码po出来。嗯🤔，仍是要坚持定时写点东西。