SIFT_DIP.py

#!/usr/bin/env python
# coding: utf-8

import cv2
import numpy as np
import os
import matplotlib.pyplot as plt


# 安装opencv库，用pip install opencv-python 会导致出错，必须指定版本
# pip install opencv-python==3.4.2.16
# pip install opencv-contrib-python==3.4.2.16
# conda install -c menpo opencv #For Anaconda User just this instead of pip

def get_train_and_test_img_features():
    train_path = "./Palmprint/training/"
    test_path = "./Palmprint/testing/"
    train_dataset = []  # 存储训练集中每张图片的SIFT特征描述向量
    test_dataset = []  # 存储测试集中每张图片的SIFT特征描述向量

    train_img_list = os.listdir(train_path)
    test_img_list = os.listdir(test_path)

    for train_img in train_img_list:
        img = cv2.imread(train_path + train_img)
        gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
        equalize = cv2.equalizeHist(gray)  # step1.预处理图片，灰度均衡化
        kp_query, des_query = get_sift_features(equalize)  # step2.获取SIFT算法生成的关键点kp和描述符des(特征描述向量)
        train_dataset.append(des_query)
    for test_img in test_img_list:
        img = cv2.imread(test_path + test_img)
        gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
        equalize = cv2.equalizeHist(gray)
        kp_query, des_query = get_sift_features(equalize)
        test_dataset.append(des_query)
    return train_dataset, test_dataset


def get_sift_features(img, dect_type='sift'):
    if dect_type == 'sift':
        sift = cv2.xfeatures2d.SIFT_create()
    elif dect_type == 'surf':
        sift = cv2.xfeatures2d.SURF_create()
    kp, des = sift.detectAndCompute(img, None)  # kp为关键点，des为描述符
    return kp, des


def sift_detect_match_num(des_query, des_train, ratio=0.70):
    # step3.使用KNN计算查询图像与训练图像之间匹配的点数目,采用k(k=2)近邻匹配，最近的点距离与次近点距离之比小于阈值ratio就认为是成功匹配。
    bf = cv2.BFMatcher()
    matches = bf.knnMatch(des_query, des_train, k=2)
    match_num = 0
    for first, second in matches:
        if first.distance < ratio * second.distance:
            match_num = match_num + 1
    return match_num


def get_one_palm_match_num(des_query, index, train_dataset, ratio=0.70):
    # 获取查询图像与训练图像中属于每一组的3张图像的匹配点的数量和
    match_num_sum = 0
    for i in range(index, index + 3):
        match_num_sum += sift_detect_match_num(des_query, train_dataset[i], ratio=ratio)
    return match_num_sum


def get_match_result(des_query, train_dataset, ratio=0.70):
    # step4.根据最大的匹配点数量和，确定查询图片的类别
    index = 0
    train_length = len(train_dataset)
    result = np.zeros(train_length // 3, dtype=np.int32)
    while index < train_length:
        result[index // 3] = get_one_palm_match_num(des_query, index, train_dataset, ratio=ratio)
        index += 3
    return result.argmax()


def predict(train_features, test_features, ratio=0.70):
    predict_true = 0
    for i, feature in enumerate(test_features):
        print('Processing image', i + 1,'...')
        # 预测每张测试图片的类别
        category = get_match_result(feature, train_features, ratio=ratio)
        if category == i // 3:
            predict_true += 1
        print('Predict result:', category + 1, 'Groud truth:', i // 3 + 1)
    print('Predict the correct number of pictures:', predict_true, 'Accuracy:', predict_true / len(test_features), 'ratio:', ratio)
    return predict_true / len(test_features)


def show_plot(ratio, acc, name, title):
    # 绘制准确率变化图
    plt.plot(ratio, acc)
    plt.title(title)
    if not os.path.exists('Image_result'):
        os.makedirs('Image_result')
    plt.savefig(os.path.join('Image_result', name))


def main():
    train_sift_features, test_sift_features = get_train_and_test_img_features()  # 存储每张图片的SIFT特征描述向量
    ratio = 0.65
    best_acc = 0
    best_ratio = 0
    ratio_list = []
    acc_list = []
    max_ratio = 0.85
    while ratio <= max_ratio:  # 循环测试具有最高准确率的ratio
        acc = predict(train_sift_features, test_sift_features, ratio)
        acc_list.append(acc)
        ratio_list.append(ratio)
        if acc > best_acc:
            best_acc = acc
            best_ratio = ratio
        ratio += 0.01
    title = 'best ratio:' + str(best_ratio) + " best acc:{:.4f}".format(best_acc)
    plt_name = "SIFT_" + str(max_ratio).split('.')[-1]
    show_plot(ratio_list, acc_list, plt_name, title)
    print(title)


if __name__ == '__main__':
    main()