检测框的输入时进行回缩（输入图像裁剪）

在输入图片时对图片去掉边上一圈，针对外扩图片的矫正：

img = cv2.imread(filename, cv2.IMREAD_COLOR)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

flag = True
image_h, image_w = img.shape[:2]
off_h = int(image_h / 10)
off_w = int(image_w / 10)
cur_h = image_h - off_h * 2
cur_w = image_w - off_w * 2
if flag:
   img = img[off_h: off_h + cur_h, off_w:off_w + cur_w]

PIL的图像外扩矫正：

img = Image.open(os.path.join(img_root,subdir,file))
w, h = img.size
off_h = h / 10
off_w = w / 10
cur_h = h - off_h * 2
cur_w = w - off_w * 2
crop = img.crop((off_w, off_h, cur_w, cur_h))
ph = os.path.join(img_root, "crop_images", file)
crop.save(ph)