【代码篇】图像预处理阶段：提取图像特征

前言

需要提取的图像特征：
1.文件大小；2.宽；3.高；4.图片尺寸；5.图像梯度（表征图像纹理的复杂程度）；6.sobel算子（表征图像复杂度）；7.色彩丰富度；8.无参考图像评价指标NIQE。

最后根据上述图像信息可对相同的图片进行去重操作

代码

# *encoding=utf-8
import re
import os
import cv2
import math
import numpy as np
import pandas as pd
import skvideo.measure


# sobel算子
def SI_IMG(img):
    sobelx = cv2.Sobel(img, cv2.CV_64F, dx=1, dy=0)
    sobelx = cv2.convertScaleAbs(sobelx).astype('float32')

    sobely = cv2.Sobel(img, cv2.CV_64F, dx=0, dy=1)
    sobely = cv2.convertScaleAbs(sobely).astype('float32')

    a = sobelx * sobelx + sobely * sobely
    result = np.sqrt(a).astype('float32')

    stddv = result.std()
    return stddv


# 色彩丰富度
def colorfulness_img(img):
    B, G, R = cv2.split(img)
    rg = R - G
    yb = (R + G) / 2 - B

    mean_rg = rg.mean()
    stddv_rg = rg.std()

    mean_yb = yb.mean()
    stddv_yb = yb.std()

    a1 = math.sqrt(stddv_rg * stddv_rg + stddv_yb * stddv_yb)
    a2 = 0.3 * math.sqrt(mean_yb * mean_yb + mean_rg * mean_rg)
    a = a1 + a2
    return a


# 提取图像特征
def get_img_info(img_path):
    try:
        img = cv2.imread(img_path, flags=cv2.IMREAD_COLOR)
    except:
        return None

    file_size = os.path.getsize(img_path)
    img_height, img_width = img.shape[:2]

    # img = img[img_height//4: img_height//4*3, img_width//4: img_width//4*3]
    # img_height, img_width = img.shape[:2]

    img_dy = img[:img_height - 1] - img[1:]
    img_dx = img[:, :img_width - 1] - img[:, 1:]
    img_gradient = np.mean(np.abs(img_dx)) + np.mean(np.abs(img_dy))

    img_si = SI_IMG(img)
    img_colorful = colorfulness_img(img)

    if img_height > 192 and img_width > 192:
        img_gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
        try:
            img_niqe = skvideo.measure.niqe(img_gray)[0]
        except:
            img_niqe = -1
    else:
        img_niqe = -1

    return file_size, img_height, img_width, img_height * img_width, img_gradient, img_si, img_niqe, img_colorful


if __name__ == '__main__':
    root_dir = "../Image-Downloader-master/download_images/gan/emoji_combine"  # emoji_all"
    file_suffix = "jpeg|jpg|png"
    output_excel = "img_combine_info.xls"

    img_info_list = []
    img_name_list = []
    file_list = os.listdir(root_dir)
    for img_name in file_list:
        # 对处理文件的类型进行过滤
        if re.search(file_suffix, img_name) is None:
            continue
        print(img_name)
        img_path = root_dir + "/" + img_name
        img_info = get_img_info(img_path)
        if img_info is None:
            print("error:", img_name)
            continue
        img_info_list.append(img_info)
        img_name_list.append(img_name)
    img_info_np = np.array(img_info_list)
    img_info_df = pd.DataFrame(img_info_np, index=img_name_list,
                               columns=["size", "height", "width", "area", "gradient", "si", "niqe", "colorful"])
                               
    # 去重
    img_info_df = img_info_df.drop_duplicates()

    # 输出统计特征
    writer = pd.ExcelWriter(output_excel)
    img_info_df.to_excel(writer, index=True, float_format="%.f")
    writer.save()

结果展示