OpenCV-iOS计算图片相似度
简书更新停留在17年?????
-
需求分析:
1、寻宝活动,商家藏宝并将场景拍照上传至服务器,用户根据线索到达指定地点,打开app进行实时扫描,如扫描到的图片与服务器的图片匹配成功,则视为中奖;
2、OpenCV中用于图片对比及识别的算法较多,具体原理及优缺点本文不做介绍,demo中用到的是特征点对比中的ORB算法;
3、本次需求中用到两个算法:特征点-ORB和直方图;
经反复测试发现特征点对比较为适合棱角分明的场景,比如桌子/电脑等;而直方图对比较为适合复杂且棱角较少的场景,比如办公室的盆栽类;(ps:测试分为室内测试及室外测试,且测试次数及测试量都不太大,结论仅限于参考)
-
OpenCV集成
1、OpenCV下载:官网地址
2、demo中OpenCV版本为3.4.3;
3、添加依赖库:
OpenCV依赖库
-
功能开发:
1、拍照上传:
1)将.m文件后缀改为.mm;
2)导入头文件(务必放在所有头文件之上)
#import <opencv2/opencv.hpp>
#import <opencv2/imgproc/types_c.h>
#import <opencv2/videoio/cap_ios.h>
#import <opencv2/imgcodecs/ios.h>
3)打开相机具体实现:
@interface CaptureView()<CvVideoCameraDelegate>
@property (strong, nonatomic) UIImageView *CvImageV;
@property (nonatomic,retain) CvVideoCamera *videoCamera;
@end
@implementation CaptureView
- (instancetype)initWithFrame:(CGRect)frame{
self = [super initWithFrame:frame];
if (self) {
[self setupCvVideoCamera];
}
return self;
}
// MARK: - init
-(void)setupCvVideoCamera{
self.CvImageV = [[UIImageView alloc]initWithFrame:CGRectMake(0, 0, self.frame.size.width, self.frame.size.height)];
self.videoCamera = [[CvVideoCamera alloc]initWithParentView:self.CvImageV];
self.videoCamera.delegate = self;
self.videoCamera.defaultAVCaptureDevicePosition = AVCaptureDevicePositionBack;
self.videoCamera.defaultAVCaptureSessionPreset = AVCaptureSessionPresetiFrame960x540;
self.videoCamera.defaultAVCaptureVideoOrientation = AVCaptureVideoOrientationPortrait;
self.videoCamera.grayscaleMode = NO;
self.videoCamera.defaultFPS = 30;
[self addSubview:self.CvImageV];
}
// MARK: - CvVideoCameraDelegate
- (void)processImage:(cv::Mat &)image{
// 将Mat转换为Xcode的UIImageView显示
UIImage *currentImage = MatToUIImage(image);
}
// MARK: - event and response
// 开启
- (void)start{
[self.videoCamera start];
}
// 暂停
- (void)stop{
[self.videoCamera stop];
}
4)问题:opencv返回的图片矩阵为BGR,需要转成RGB来上传,具体代码见demo
2、图片对比:
1)将.m文件后缀改为.mm;
2)导入头文件(务必放在所有头文件之上)
#import <opencv2/opencv.hpp>
#import <opencv2/imgcodecs/ios.h>
#include "opencv2/core/core.hpp"
#include <iostream>
#include <vector>
3)实现代码:
/**
@param boxImage 模板图片
@param senceImage 实时图片
@return 对比结果(见图)
*/
-(UIImage *)similarlyMatchWithBox:(UIImage *)boxImage andSence:(UIImage *)senceImage{
if (boxImage && senceImage) {
cv::Mat sence,box;
box = [self cvMatFromUIImage:boxImage];
sence = [self cvMatFromUIImage:senceImage];
cvtColor(box, box, CV_RGBA2RGB);
cvtColor(sence, sence, CV_RGBA2RGB);
vector<KeyPoint> keyPoints_obj, keyPoints_sence;
Ptr<ORB> detector = ORB::create();
detector->detect(box, keyPoints_obj);
detector->detect(sence, keyPoints_sence);
Mat description_box,description_sence;
detector->compute(box, keyPoints_obj, description_box);
detector->compute(sence, keyPoints_sence, description_sence);
vector<DMatch> matches;
Ptr<DescriptorMatcher> matcher = DescriptorMatcher ::create(cv::BFMatcher::BRUTEFORCE);
// 处理拍的区域为黑色,比如手机平放在桌面上,会闪退
if (description_sence.cols <= 0) {
return nil;
}
if (description_sence.rows <= 0) {
return nil;
}
matcher->match(description_box, description_sence, matches);
// 发现匹配
vector<DMatch> good_matches;
for (unsigned int i = 0; i < matches.size(); i++) {
//distance 值根据实际开发来调试
if (matches[i].distance <= 320) {
good_matches.push_back(matches[i]);
}
}
Mat imgMatches;
drawMatches(box,keyPoints_obj,sence,keyPoints_sence,good_matches,imgMatches,Scalar::all(-1),Scalar::all(-1),vector<char>(),DrawMatchesFlags::NOT_DRAW_SINGLE_POINTS);
UIImage *m = [self UIImageFromCVMat:imgMatches];
return m;
}
return nil;
}
4)具体代码见demo中ImageCompared
文件;
5)问题:
① 相机拍到黑色时会崩溃:可通过description_sence.cols <= 0 && description_sence.rows <= 0
来判断;这个坑我和安卓都爬了好久才爬出来;(悄悄告诉你我先爬出来的,美滋滋)
② 存储在服务器的模板图可多上传几张,一张图片匹配成功率会很低:可设置至少上传X张,且在上传的时候也进行匹配,匹配率达到某个值则可上传,否则要求商家重拍,防止上传的图片不是同一个场景;
③ 内存问题:实时扫描且进行相似度对比消耗内存非常大,很快就会内存溢出导致闪退,我的解决方法是设置为每隔X秒进行一次匹配,时间可自己调试然后给定;
6)匹配结果图:
匹配结果
- 最后:OpenCV里的学问太多了,这次需求甲方爸爸赶得急是一方面,之前没了解过也是真的,熬了几个通宵才搞定,大概算是一只脚迈进去了,如果二期需求有时间排的话可能另一只脚也就能迈进去了,如果没有的话就此打住。