图像分类学习笔记DAY03

1.蝴蝶图像分类
思路一:数据增强

图像增强之numpy格式
1）图像缩放

def resize_img(img, target_size):
    img = cv2.resize(img, (target_size, target_size))
    return img

2）图片中央剪切

def center_crop_img(img, target_size):
    w, h = img.shape[0], img.shape[1]
    tw, th = target_size, target_size
    assert (w >= target_size) and (h >= target_size), \
            "image width({}) and height({}) should be larger than crop size".format(w, h, target_size)
    x1 = int(round((w - tw) / 2.))
    y1 = int(round((h - th) / 2.))
    img = img[x1:x1+tw, y1:y1+th]
    return img

3）随机剪切

def random_crop_img(img, target_size, scale=[0.08, 1.0], ratio=[3. / 4., 4. / 3.]):
    aspect_ratio = math.sqrt(np.random.uniform(*ratio))
    w = 1. * aspect_ratio
    h = 1. / aspect_ratio

    bound = min((float(img.shape[1]) / img.shape[0]) / (w**2),
                (float(img.shape[0]) / img.shape[1]) / (h**2))
    scale_max = min(scale[1], bound)
    scale_min = min(scale[0], bound)

    target_area = img.shape[1] * img.shape[0] * np.random.uniform(scale_min,
                                                                scale_max)
    target_size = math.sqrt(target_area)
    w = int(target_size * w)
    h = int(target_size * h)

    i = np.random.randint(0, img.shape[1] - w + 1)
    j = np.random.randint(0, img.shape[0] - h + 1)

    img = img[i:i+w, j:j+h]
    img = cv2.resize(img, (int(target_size), int(target_size)))
    return img

4)随机旋转

def rotate_image(img):
    # 将图片随机旋转-14到15之间的某一个角度
    angle = np.random.randint(-14, 15)
    #获取图像的尺寸
    #旋转中心
    h, w = img.shape[0], img.shape[1]
    cx,cy = w/2,h/2
    #设置旋转矩阵
    M = cv2.getRotationMatrix2D((cx,cy),-angle,1.0)
    cos = np.abs(M[0,0])
    sin = np.abs(M[0,1])
    # 计算图像旋转后的新边界
    nW = int((h*sin)+(w*cos))
    nH = int((h*cos)+(w*sin))
    # 调整旋转矩阵的移动距离（t_{x}, t_{y}）
    M[0,2] += (nW/2) - cx
    M[1,2] += (nH/2) - cy
    return cv2.warpAffine(img,M,(nW,nH))

5）随机左右翻转

def flip_image(img):
    # 将图片随机左右翻转， 根据需要也可以设置随机上下翻转
    v = random.random()
    if v < 0.5:
        img = cv2.flip(img, 1)
    return img

6）亮度，饱和度，色度随机调整

def distort_image(img):
    # 随机数据增强
    v = random.random()
    # 顺序可以自己随意调整
    if v < 0.35:
        img = bright_image(img)
        img = saturation_image(img)
        img = hue_image(img)
    elif v < 0.7:
        img = bright_image(img)
        img = saturation_image(img)
        img = hue_image(img)
    return img

7）Cutout
Cutout 可以理解为 Dropout 的一种扩展操作，不同的是 Dropout 是对图像经过网络后生成的特征进行遮挡，而 Cutout 是直接对输入的图像进行遮挡，相对于Dropout对噪声的鲁棒性更好。PaddleClas中已有该方法的实现。这里将其改写为Paddle2.0中的自定义数据增强类。

import numpy as np

class Cutout(paddle.vision.transforms.BaseTransform):
    def __init__(self, n_holes=1, length=112, prob=0.5, keys=None):
        super(Cutout, self).__init__(keys)
        self.prob = prob
        self.n_holes = n_holes
        self.length = length

    def _get_params(self, inputs):
        image = inputs[self.keys.index('image')]
        params = {}
        params['cutout'] = np.random.random() < self.prob
        # params['size'] = _get_image_size(image)
        return params

    def _apply_image(self, img):
        """ cutout_image """
        if self.params['cutout']:
            h, w = img.shape[:2] ## input image, with shape of (H, W, C) 
            mask = np.ones((h, w), np.float32)

            for n in range(self.n_holes):
                y = np.random.randint(h)
                x = np.random.randint(w)

                y1 = np.clip(y - self.length // 2, 0, h)
                y2 = np.clip(y + self.length // 2, 0, h)
                x1 = np.clip(x - self.length // 2, 0, w)
                x2 = np.clip(x + self.length // 2, 0, w)

                img[y1:y2, x1:x2] = 0
            return img
        else:
            return img

思路二：优化模型
尝试换了模型，模型能力太强，有时会导致预测准确率不高。

小结：通过思路一的数据增强方法后，将蝴蝶图像分类准确率提高到了85%以上。若想将准确率提高至90%须考虑优化模型。

蝴蝶图像分类项目：https://aistudio.baidu.com/aistudio/projectdetail/1618006

参考文档：https://aistudio.baidu.com/aistudio/projectdetail/699878
https://aistudio.baidu.com/aistudio/projectdetail/1357939