# -*- coding: utf-8 -*-
"""
File Name： image
Description : 图像处理工具类
Author : mick.yi
date： 2019/2/18
"""
import skimage
from skimage import io, transform
import numpy as np
import matplotlib.pyplot as plt
import random
def load_image(image_path):
"""
加载图像
:param image_path: 图像路径
:return: [h,w,3] numpy数组
"""
image = plt.imread(image_path)
# 灰度图转为RGB
if len(image.shape) == 2:
image = np.expand_dims(image, axis=2)
image = np.tile(image, (1, 1, 3))
elif image.shape[-1] == 1:
image = skimage.color.gray2rgb(image) # io.imread 报ValueError: Input image expected to be RGB, RGBA or gray
# 标准化为0~255之间
if image.dtype == np.float32:
image *= 255
image = image.astype(np.uint8)
# 删除alpha通道
return image[..., :3]
def load_image_gt(image_id, image_path, output_size, gt_boxes=None,
gt_quadrilaterals=None, horizontal_flip=False, random_crop=False):
"""
加载图像，生成训练输入大小的图像，并调整GT 边框，返回相关元数据信息
:param image_id: 图像编号id
:param image_path: 图像路径
:param output_size: 图像输出尺寸，及网络输入到高度或宽度(默认长宽相等)
:param gt_boxes: GT 边框 [N,(y1,x1,y2,x2)]
:param gt_quadrilaterals:
:param horizontal_flip: 是否水平翻转
:param random_crop: 是否随机裁剪
:return:
image: (H,W,3)
image_meta: 元数据信息，详见compose_image_meta
gt_boxes：图像缩放及padding后对于的GT 边框坐标 [N,(y1,x1,y2,x2)]
"""
# 加载图像
image = load_image(image_path)
# 随机裁剪
if random_crop and random.random() >= 0.5:
min_x, max_x = np.min(gt_quadrilaterals[:, ::2]), np.max(gt_quadrilaterals[:, ::2])
min_y, max_y = np.min(gt_quadrilaterals[:, 1::2]), np.max(gt_quadrilaterals[:, 1::2])
image, crop_window = crop_image(image, [min_y, min_x, max_y, max_x])
# print(image.shape,[min_y, min_x, max_y, max_x],crop_window)
# gt坐标偏移
if gt_quadrilaterals is not None and gt_quadrilaterals.shape[0] > 0:
gt_quadrilaterals[:, 1::2] -= crop_window[0]
gt_quadrilaterals[:, ::2] -= crop_window[1]
# 水平翻转
if horizontal_flip:
image = image[:, ::-1, :]
# gt翻转
if gt_quadrilaterals is not None and gt_quadrilaterals.shape[0] > 0:
gt_quadrilaterals[:, ::2] = image.shape[1] - gt_quadrilaterals[:, ::2]
lt_x, lt_y, rt_x, rt_y, rb_x, rb_y, lb_x, lb_y = np.split(gt_quadrilaterals, 8, axis=1)
gt_quadrilaterals = np.concatenate([rt_x, rt_y, lt_x, lt_y, lb_x, lb_y, rb_x, rb_y], axis=1)
original_shape = image.shape
# resize图像，并获取相关元数据信息
image, window, scale, padding = resize_image(image, output_size)
# 组合元数据信息
image_meta = compose_image_meta(image_id, original_shape, image.shape,
window, scale)
# 根据缩放及padding调整GT边框
if gt_boxes is not None and gt_boxes.shape[0] > 0:
gt_boxes = adjust_box(gt_boxes, padding, scale)
if gt_quadrilaterals is not None and gt_quadrilaterals.shape[0] > 0:
gt_quadrilaterals = adjust_quadrilaterals(gt_quadrilaterals, padding, scale)
return image, image_meta, gt_boxes, gt_quadrilaterals
def crop_image(image, gt_window):
h, w = list(image.shape)[:2]
y1, x1, y2, x2 = gt_window
gaps = np.array([y1, x1, h - y2, w - x2])
wy1 = np.random.randint(min(y1+1, h // 5))
wx1 = np.random.randint(min(x1+1, w // 5))
wy2 = h - np.random.randint(min(h - y2 + 1, h // 5))
wx2 = w - np.random.randint(min(w - x2 + 1, w // 5))
return image[wy1:wy2, wx1:wx2], [wy1, wx1, wy2, wx2]
def resize_image(image, max_dim):
"""
缩放图像为正方形，指定长边大小，短边padding;
:param image: numpy 数组(H,W,3)
:param max_dim: 长边大小
:return: 缩放后的图像,元素图像的宽口位置，缩放尺寸，padding
"""
image_dtype = image.dtype
h, w = image.shape[:2]
scale = max_dim / max(h, w) # 缩放尺寸
image = transform.resize(image, (round(h * scale), round(w * scale)),
order=1, mode='constant', cval=0, clip=True, preserve_range=True)
h, w = image.shape[:2]
# 计算padding
top_pad = (max_dim - h) // 2
bottom_pad = max_dim - h - top_pad
left_pad = (max_dim - w) // 2
right_pad = max_dim - w - left_pad
padding = [(top_pad, bottom_pad), (left_pad, right_pad), (0, 0)]
image = np.pad(image, padding, mode='constant', constant_values=0)
# 原始图像在缩放图像上的窗口位置
window = (top_pad, left_pad, h + top_pad, w + left_pad) #
return image.astype(image_dtype), window, scale, padding
def compose_image_meta(image_id, original_image_shape, image_shape,
window, scale):
"""
组合图像元数据信息，返回numpy数据
:param image_id:
:param original_image_shape: 原始图像形状，tuple(H,W,3)
:param image_shape: 缩放后图像形状tuple(H,W,3)
:param window: 原始图像在缩放图像上的窗口位置（y1,x1,y2,x2)
:param scale: 缩放因子
:return:
"""
meta = np.array(
[image_id] + # size=1
list(original_image_shape) + # size=3
list(image_shape) + # size=3
list(window) + # size=4 (y1, x1, y2, x2) in image cooredinates
[scale] # size=1
)
return meta
def parse_image_meta(meta):
"""
解析图像元数据信息,注意输入是元数据信息数组
:param meta: [12]
:return:
"""
image_id = meta[0]
original_image_shape = meta[1:4]
image_shape = meta[4:7]
window = meta[7:11] # (y1, x1, y2, x2) window of image in in pixels
scale = meta[11]
return {
"image_id": image_id.astype(np.int32),
"original_image_shape": original_image_shape.astype(np.int32),
"image_shape": image_shape.astype(np.int32),
"window": window.astype(np.int32),
"scale": scale.astype(np.float32)
}
def batch_parse_image_meta(meta):
"""
解析图像元数据信息,注意输入是元数据信息数组
:param meta: [batch,12]
:return:
"""
image_id = meta[:, 0]
original_image_shape = meta[:, 1:4]
image_shape = meta[:, 4:7]
window = meta[:, 7:11] # (y1, x1, y2, x2) window of image in in pixels
scale = meta[:, 11]
return {
"image_id": image_id.astype(np.int32),
"original_image_shape": original_image_shape.astype(np.int32),
"image_shape": image_shape.astype(np.int32),
"window": window.astype(np.int32),
"scale": scale.astype(np.float32)
}
def adjust_box(boxes, padding, scale):
"""
根据填充和缩放因子，调整boxes的值
:param boxes: numpy 数组; GT boxes [N,(y1,x1,y2,x2)]
:param padding: [(top_pad, bottom_pad), (left_pad, right_pad), (0, 0)]
:param scale: 缩放因子
:return:
"""
boxes = boxes * scale
boxes[:, 0::2] += padding[0][0] # 高度padding
boxes[:, 1::2] += padding[1][0] # 宽度padding
return boxes
def adjust_quadrilaterals(quadrilaterals, padding, scale):
"""
根据填充和缩放因子，调整四边形的值
:param quadrilaterals: numpy 数组; GT quadrilaterals[N,(x1,y1,x2,y2,x3,y3,x4,y4)]
:param padding: [(top_pad, bottom_pad), (left_pad, right_pad), (0, 0)]
:param scale: 缩放因子
:return:
"""
quadrilaterals = quadrilaterals * scale
quadrilaterals[:, 1::2] += padding[0][0] # 高度padding
quadrilaterals[:, 0::2] += padding[1][0] # 宽度padding
return quadrilaterals
def recover_detect_boxes(boxes, window, scale):
"""
将检测边框映射到原始图像上，去除padding和缩放
:param boxes: numpy数组，[n,(y1,x1,y2,x2)]
:param window: [(y1,x1,y2,x2)]
:param scale: 标量
:return:
"""
# 去除padding
boxes[:, 0::2] -= window[0]
boxes[:, 1::2] -= window[1]
# 还原缩放
boxes /= scale
return boxes
def recover_detect_quad(boxes, window, scale):
"""
将检测四边形映射到原始图像上，去除padding和缩放
:param boxes: numpy数组，[n,(x1,y1,x2,y2,x3,y3,x4,y4)]
:param window: [(y1,x1,y2,x2)]
:param scale: 标量
:return:
"""
# 去除padding
boxes[:, 1::2] -= window[0] # 高度
boxes[:, 0::2] -= window[1] # 宽度
# 还原缩放
boxes /= scale
return boxes
