[PYTHON] Cat detection with OpenCV (model distribution)

(Added on 2014/09)

I also write an article on how to identify cat breeds using deep learning techniques. Please refer if you are interested.

• Cat breed identification by deep learning

Cat detection with OpenCV

(Chocolat-Kichijoji Calico)

Cat image processing part 1 – Collecting materials organized how to collect cat images, but here we use the collected cat images to organize cats. I created a learning model (classifier) for detection.

If you are interested, the technical details are written on the blog. Model distribution will be discussed later in this entry.

• Cat and image processing part 2-Cat detection

Learning is with the familiar Boosting (one of the ensemble learning methods). For the features, we adopted LBP features, which can quickly repeat model creation and verification. In addition, about 7,000 annotation data were collected using crowdsourcing.

How to use

C++ From C ++, use it as follows.

// objdetect.cpp
#include <iostream>
#include <vector>
#include <opencv2/opencv.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/objdetect/objdetect.hpp>

int main(int argc, char** argv) {
  using namespace std;
  if(argc != 3) {
    cerr << "./objdetect [image file] [cascade file]" << endl;
    exit(-1);
  }
  cv::Mat src_img = cv::imread(argv[1], 1);
  if(src_img.empty()) {
    cerr << "cannot load image" << endl;
    exit(-1);
  }
  cv::Mat dst_img = src_img.clone();
  string cascade_file = string(argv[2]);
  cv::CascadeClassifier cascade;
  cascade.load(cascade_file);
  if(cascade.empty()) {
    cerr << "cannot load cascade file" << endl;
    exit(-1);
  }
  vector<cv::Rect> objects;
  cascade.detectMultiScale(src_img, objects, 1.1, 3);
  vector<cv::Rect>::const_iterator iter = objects.begin();
  cout << "count: " << objects.size() << endl;
  while(iter!=objects.end()) {
    cout << "(x, y, width, height) = (" << iter->x << ", " << iter->y << ", "
         << iter->width << ", " << iter->height << ")" << endl;
    cv::rectangle(dst_img,
                  cv::Rect(iter->x, iter->y, iter->width, iter->height),
                  cv::Scalar(0, 0, 255), 2);
    ++iter;
  }
  cv::imwrite("result.jpg ", dst_img);
  return 0;
}

##Compile and run
$ g++ -o objdetect objdetect.cpp -I/path/to/opencv/include -L/path/to/opencv/lib -lopencv_core -lopencv_highgui  -lopencv_objdetect
$ ./objdetect cat.jpg cascade.xml
count: 1
(x, y, width, height) = (278, 185, 65, 65)

It seems that the detection accuracy will be improved if the parameters at the time of detection are specified loosely (minNeighbors parameter is set to a small value). Of course, false positives will increase, but if you add processing such as excluding the case where the detection area is smaller than a certain size (specifying the minSize parameter), you can make it look like a certain degree of accuracy.

CascadeClassifier::detectMultiScale - docs.opencv.org

Python Easy to use from Python.

#!/usr/local/bin/python
# -*- coding: utf-8 -*-

import sys
import cv2 as cv

def detect(imagefilename, cascadefilename):
    srcimg = cv.imread(imagefilename)
    if srcimg is None:
        print('cannot load image')
        sys.exit(-1)
    dstimg = srcimg.copy()
    cascade = cv.CascadeClassifier(cascadefilename)
    if cascade.empty():
        print('cannnot load cascade file')
        sys.exit(-1)
    objects = cascade.detectMultiScale(srcimg, 1.1, 3)
    for (x, y, w, h) in objects:
        print(x, y, w, h)
        cv.rectangle(dstimg, (x, y), (x + w, y + h), (0, 0, 255), 2)
    return dstimg

if __name__ == '__main__':
    result = detect('cat.jpg', 'cascade.xml')
    cv.imwrite('result.jpg', result)

Model distribution

I will publish the model I made this time. It is a cascade file (cascade.xml) that can be read by OpenCV.

• cat detection cascade model - wellflat/cat-fancier

Since we are running a batch process on our personal server that detects the new addition of teacher data to the database and runs learning, it should be continuously and automatically updated to a more accurate model.

That said, if no measures are taken, the accuracy will reach a plateau, so I think that the algorithm will be improved by adding various pre-processing and post-processing.

• Cat and image processing part 1 – Collecting materials
• Cat and image processing part 2-Cat detection