-1

I want to analyze my own images using an SVM trained on the MNIST dataset. How can I preprocessed my image so it can be accepted by the model?

dataset = datasets.fetch_openml("mnist_784", version=1)
(trainX, testX, trainY, testY) = train_test_split(
    dataset.data / 255.0, dataset.target.astype("int0"), test_size = 0.33)

ap = argparse.ArgumentParser()
ap.add_argument("-d", "--dataset", type=str, default="3scenes",
    help="path to directory containing the '3scenes' dataset")
ap.add_argument("-m", "--model", type=str, default="knn",
    help="type of python machine learning model to use")

args = vars(ap.parse_args())

#user input image to classify

userImage = cv.imread('path_to_image/1.jpg')

#preprocess user image
#...

models = {
    "svm": SVC(kernel="linear"),
}

# train the model
print("[INFO] using '{}' model".format(args["model"]))
model = models[args["model"]]
model.fit(trainX, trainY)

print("[INFO] evaluating image...")
predictions = model.predict(userImage)
print(classification_report(userImage, predictions))
Carlos Hevia
  • 11
  • 6
  • What exactly all this code is about? – desertnaut Oct 22 '19 at 20:34
  • Hard to tell without seeing those images. – T A Oct 23 '19 at 14:40
  • I just want to classify my image using a model trained on the MNIST dataset. My question was how to transform mi image to the same format of the MNIST dataset so I can pass it through the model and get a classification. Maybe that's not the way it works, I am a beginner at this topic, please let me know any insight. – Carlos Hevia Oct 25 '19 at 03:14

1 Answers1

0

MNIST images have the following shape: 28x28x1, width 28 pixels, height 28 pixels and one color channel i.e. grayscale.

Assuming your model takes the same input shape, you can use the following:

import cv2
userImage = cv2.imread('path_to_image/1.jpg')
# resize image to 28x28
userImage = cv2.resize(userImage,(28,28))
# convert to grayscale
userImage = cv2.cvtColor(userImage,cv2.COLOR_BGR2GRAY)
# normalize
userImage /= 255.

Depending on how large your image is, you may want to select an 28x28 patch manually. Otherwise you risk of losing image quality and thus information.

If you model takes a vector as input, you can use the following to flatten your image before feeding it to the model:

userImage = np.reshape(userImage,(784,))
Tinu
  • 2,432
  • 2
  • 8
  • 20