part3.md

Part 3: Say Cheese!

Files: neg.py, create_samples.py

result:

Creating Dataset

For haar cascade check out this tutorial

learning a gray image is faster than an image with 3 channels

Well for training we require background samples which will act as negative image samples, so grayscale all gen_images and add their path to bg.txt also when you check page source of Captcha of kissanime, there exists a attribute 'width=160', so we need to resize all images hence the following:

h,w = img.shape
h = int(numpy.round((h*160/w),0)) ##rounding off to integer(no decimal place)
w = 160

neg.py

Now the background images are done time to add number in there foreground.

for that take a look at this tutorial.

We take a number from 'num_imgs' folder randomly resize it with a max factor of 0.2 and randomly place somewhere onto background image (with some margin).

number = cv2.imread('num_imgs/<img>')
scaling_factor = numpy.random.randint(8,11)

h,w,_ = number.shape ##channels not required
w_0 = int(numpy.round((scaling_factor*w/10),0))
h_0 = int(numpy.round((scaling_factor*h/10),0))

create_samples.py

For the code look at create_samples.py.

Training Model: Command line stuff

Now create a positive.vec

opencv_createsamples -info info.txt -num <number_of_samples_genrated> -vec positives.vec -h 55 -w 45

and then train it! (make sure haar_cascade folder do exists)

opencv_traincascade -data haar_cascade -vec positives.vec -bg bg.txt -numPos 16000 -numNeg 8000 -numStages 16 -h 55 -w 45

number of stages that I used was 16 (took 3.5 day on WSL), you can use more if you train it till 17 then first 16 will be loaded from xml files present in 'haar_cascade' folder

And the Haar Cascade is trained!

Now one can use .xml or convert .xml to .js (as shown in last Part) for extension

Previous Part

Next Part