比较使用 ImageDataGenerator() 和 cv2.imread() 加载的数据

Question

I am new to image classification. I have built a model to classify dogs and cats and saved the model as an h5 file. For training, data was loaded using ImageDataGenerator() . For testing, I used cv2.imread() to load the data. For confirmation, I just loaded an image using both these methods and checked the output. But the array I got was reversed from what I got from ImageDataGenerator() . I am posting the code and output

train_image_generator = ImageDataGenerator()
test_data=train_image_generator.flow_from_directory(batch_size=batch_size,directory='/home/josin/my_projects/test1',shuffle=True,target_size=(150, 150),class_mode='binary')
print(test_data[0][0])`

The output for the above code is

Found 1 images belonging to 1 classes

array([[[[203., 164.,  87.]
     [209., 170.,  93.],
     [209., 170.,  93.],
     ...,
     [247., 206., 124.],
     [244., 204., 119.],
     [240., 201., 122.]],
     ...,
     [  2.,   2.,   0.],
     [  2.,   2.,   0.],
     [  2.,   2.,   0.]]]], dtype=float32)`

The code using `cv2.imread() is:

img_array = cv2.imread('/home/josin/my_projects/test1/cat/cat.0.jpg')
new_array = cv2.resize(img_array,(150,150))
new_array.reshape(1,150,150,3)`

and output of the above code is

array([[[[ 87, 164, 203],
     [ 92, 169, 208],
     [ 93, 170, 209],
     ...,
     [124, 206, 247],
     [119, 203, 245],
     [122, 201, 240]],
     ...,
     [  0,   2,   2],
     [  0,   2,   2],
     [  0,   2,   2]]]], dtype=uint8)

The pixel values got reversed while using cv2.imread() .Most of the articles and posts I have referred uses cv2.imread() .Is it correct to give this input to the saved model? Is there any other way to load images to serve to model

Answer 1

cv2.imread arranges the image channels as follows: b, g, r . Meanwhile, ImageDataGenerator arranges channels as r, g, b . This explains the inverse effect you mentioned.

Simply, call: img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) to make opencv images arranged as r, g, b