FAST角點檢測算法（二）- 非極大值抑制篩選fast特徵點

FAST角點檢測算法（二）- 非極大值抑制篩選fast特徵點

---

author@jason_ql（lql0716）
http://blog.csdn.net/lql0716

---

• fast角點檢測算法參考文章《fast角點檢測算法》（涵蓋fast角點檢測原理及C++、python代碼，以及效果圖）

• 非極大值抑制，就是對於一個3*3（或5*5,7*7等奇數窗口）的窗口，如果存在多個特徵點，則刪除響應值較小的特徵點，只保留響應值最大的特徵點。

• 這裏根據fast特徵點的響應值的大小，做了非極大值抑制處理，對特徵點進行了篩選。

1.1 C++代碼

#include <opencv2\opencv.hpp>

using namespace cv;
using namespace std;

//局部極大值抑制，這裏利用fast特徵點的響應值做比較
void selectMax(int window, cv::Mat gray, std::vector<KeyPoint> & kp){

    //window是局部極大值抑制的窗口大小，r爲半徑
    int r = window / 2;
    if (window != 0){
        //對kp中的點進行局部極大值篩選
        for (int i = 0; i < kp.size(); i++){
            for (int j = i + 1; j < kp.size(); j++){
                //如果兩個點的距離小於半徑r，則刪除其中響應值較小的點
                if (abs(kp[i].pt.x - kp[j].pt.x) + abs(kp[i].pt.y - kp[j].pt.y) <= 2 * r){
                    if (kp[i].response < kp[j].response){
                        std::vector<KeyPoint>::iterator it = kp.begin() + i;
                        kp.erase(it);
                        selectMax(window, gray, kp);
                    }
                    else{
                        std::vector<KeyPoint>::iterator it = kp.begin() + j;
                        kp.erase(it);
                        selectMax(window, gray, kp);
                    }
                }
            }
        }
    }

}

//
void fastpoint(cv::Mat gray, int threshold, int window, int pointNum, std::vector<KeyPoint> & kp){

    std::vector<KeyPoint> keypoint;

    cv::FastFeatureDetector fast(threshold, true);  //threshold 爲閾值，越大，特徵點越少
    fast.detect(gray, keypoint);  //fast特徵檢測
    if (keypoint.size() > pointNum){
        threshold = threshold + 5;
        fastpoint(gray, threshold, window, pointNum, keypoint);
    }
    selectMax(window, gray, keypoint);
    kp.assign(keypoint.begin(), keypoint.end());    //複製可以point到kp

}


int main(){
    cv::Mat img = cv::imread("D:/photo/06.jpg");
    cv::Mat gray;
    cv::cvtColor(img, gray, cv::COLOR_BGR2GRAY);
    std::vector<KeyPoint> kp;
    int threshold = 30;  //fast閾值
    int window1 = 7;  //局部非極大值抑制窗口
    int pointMaxNum1 = 400; //特徵點最大個數
    fastpoint(gray, threshold, window1, pointMaxNum1, kp);

    cv::Mat img2, gray2;
    std::vector<KeyPoint> kp2;
    img.copyTo(img2);
    gray.copyTo(gray2);
    int window2 = 15;  //局部非極大值抑制窗口
    int pointMaxNum2 = 400; //特徵點最大個數
    fastpoint(gray2, threshold, window2, pointMaxNum2, kp2);

    cv::drawKeypoints(img, kp, img, Scalar(0, 0, 255));
    cv::drawKeypoints(img2, kp2, img2, Scalar(255, 0, 0));

    cv::imwrite("D:/photo/06_1.jpg", img);
    cv::namedWindow("img", cv::WINDOW_NORMAL);
    cv::imshow("img", img);

    cv::imwrite("D:/photo/06_2.jpg", img2);
    cv::namedWindow("img2", cv::WINDOW_NORMAL);
    cv::imshow("img2", img2);

    cv::waitKey(0);

    system("pause");
    return 0;
}

原圖：

效果圖img：

效果圖img2：

1.2 python代碼

# -*- coding: utf-8 -*-
"""
Created on Fri Jun 16 09:55:41 2017

@author: User
"""

import cv2
import numpy as np
import copy

def selectMax(window, gray, kp):
    r = window / 2
    a = 0
    if window != 0:
        for i in range(np.array(kp).shape[0]):
            for j in range(i+1, np.array(kp).shape[0]):
                if np.abs(kp[i].pt[0]-kp[j].pt[0]) + np.abs(kp[i].pt[1] - kp[j].pt[1]) <= 2*r:
                    if kp[i].response < kp[j].response:
                        kp.pop(i)                      
                        a = 1
                        break
                    else:
                        kp.pop(j)
                        a = 1
                        break
        if a != 0:
            kp = selectMax(window, gray, kp)

    return kp

def fastpoint(gray, threshold, window, pointNum):
        #fast = cv2.FeatureDetector_create('FAST')
    #cv2.FAST_FEATURE_DETECTOR_TYPE_5_8
    #cv2.FAST_FEATURE_DETECTOR_TYPE_7_12
    #cv2.FAST_FEATURE_DETECTOR_TYPE_9_16

    fasts = cv2.FastFeatureDetector(threshold)
    kp = fasts.detect(gray, None)
    if np.array(kp).shape[0] > pointNum:
        threshold = threshold + 5
        kp = fastpoint(gray, threshold, window, pointNum)

    kp0 = selectMax(window, gray, kp)
    return kp0


img = cv2.imread('D:/photo/06.jpg')
gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
threshold = 20;
window1 = 7
pointMaxNum1 = 400
kp = fastpoint(gray, threshold, window1, pointMaxNum1)


img2 = copy.deepcopy(img)
gray2 = copy.deepcopy(gray)
window2 = 15
pointMaxNum2 = 400
kp2 = fastpoint(gray2, threshold, window2, pointMaxNum2)

img = cv2.drawKeypoints(img, kp, color = (0,255,255))
img2 = cv2.drawKeypoints(img2, kp2, color = (255,255,0))

cv2.imwrite('D:/photo/06_1p.jpg', img)
cv2.namedWindow('img', cv2.WINDOW_NORMAL)
cv2.imshow('img',img)

cv2.imwrite('D:/photo/06_2p.jpg', img2)
cv2.namedWindow('img2', cv2.WINDOW_NORMAL)
cv2.imshow('img2',img2)

cv2.waitKey(0)

效果圖img：

效果圖img2：

• 相關文章
  

• 機器學習、深度學習、計算機視覺、自然語言處理及應用案例——乾貨分享（持續更新……）
  鏈接：http://blog.csdn.net/lql0716/article/details/70479493
  
• FAST角點檢測算法
  鏈接：http://blog.csdn.net/lql0716/article/details/65662648
  
• Harris角點檢測(Python-OpenCV)
  鏈接：http://blog.csdn.net/lql0716/article/details/52628959