Why are the prediction probabilities so high for images with no defined class?

Question

I am trying to create a CNN that could detect the numbers in an image. For this I started working with The Street View House Numbers (SVHN) Dataset. This dataset comes with pre-processed images scaled to 32x32 digits.There are 10 classes for 10 numbers.

I trained the network and it gives a decent test accuracy of close to ~0.93. The test accuracy is also calculated on the test set which is a set of 32x32 digits.

This is all good. But the problem is the prediction probability is always one. Here is how the output of one of the class looks like:

 array([[0.0000000e+00, 0.0000000e+00, 0.0000000e+00, 0.0000000e+00,
        1.0000000e+00, 8.5623318e-24, 0.0000000e+00, 0.0000000e+00,
        0.0000000e+00, 2.4716297e-28]], dtype=float32)

As could be seen from the output of one of the example, class probability for one of the class is 1. This is fine for an image which contains the image of the desired class but probaility of 1 happens even when there no mark of number in an image. For example, the following image predicts the class 4 with probability of 1. In fact the above distribution is for the following image.

Image:

I have not been able to identify the reason for this. I am sharing the code I used to create the CNN.

val_split_length = 10623
num_train_samples = 73257
num_test_samples = 26032
total_classes = 10
model_prefix = "10c"

model = keras.Sequential()
# First Conv. Layer
model.add(keras.layers.Conv2D(filters = 96, kernel_size = (11,11), strides = (4,4), padding = "same", input_shape=(227,227,3)))
model.add(keras.layers.Activation("relu"))
model.add(keras.layers.BatchNormalization())
model.add(keras.layers.MaxPooling2D(pool_size = (3,3), strides = (2,2), padding="same"))

# ##More Conv. Layers ###

# First Fully Connected Layer
model.add(keras.layers.Flatten())
model.add(keras.layers.Dense(4096))
model.add(keras.layers.Activation("relu"))
model.add(keras.layers.Dropout(0.5))

## More Fully Connected Layers ###

# Third Fully Connected Layer
model.add(keras.layers.Dense(total_classes))
model.add(keras.layers.Activation("softmax"))



train_optimizer_adam = tf.train.AdamOptimizer(learning_rate=1e-3)
train_optimizer_rmsProp = keras.optimizers.RMSprop(lr=0.0001)

#https://keras.io/optimizers/
model.compile(loss="categorical_crossentropy", optimizer=train_optimizer_rmsProp, metrics=['accuracy'])

batch_size = 128 * 3

data_generator = keras.preprocessing.image.ImageDataGenerator(rescale = 1./255)

# https://keras.io/preprocessing/image/#flow_from_directory
train_generator = data_generator.flow_from_directory(
        'train',
        target_size=(227, 227),
        batch_size=batch_size,
        color_mode='rgb',
        class_mode='categorical',
        #save_to_dir="logs"
)

validation_generator = data_generator.flow_from_directory(
        'validation',
        target_size=(227, 227),
        batch_size=batch_size,
        color_mode='rgb',
        class_mode='categorical')


# https://keras.io/models/model/#fit_generator
history = model.fit_generator(
    train_generator, 
    validation_data = validation_generator, 
    validation_steps = math.ceil(val_split_length / batch_size),
    epochs = 5, 
    steps_per_epoch = math.ceil(num_train_samples / batch_size), 
    use_multiprocessing = True, 
    workers = 8, 
    callbacks = model_callbacks, 
    verbose = 2
)

To predict from the model above:

img = cv2.imread("image.png")
img = cv2.resize(img, (227,227))
loaded_model = keras.models.load_model("saved-model-12-0.96.hdf5")
prob = loaded_model.predict_proba(np.expand_dims(img, axis = 0))
print(prob)

What could be the reason that I get a high probability for class that does not exist anywhere in the image? I understand that the model will predict something, but why is the probability so high?

This is not really a programming question. The reason is because the probabilities are not calibrated, so they don't have any real meaning as probabilities. The model has always been trained with 1.0 probabilities, so its expected that it will perform like that. — Dr. Snoopy
@MatiasValdenegro I'd like to disagree with you. It's very unlikely that you get such a confident prediction (i.e. max prob of 1) in a CNN, even in a very good model. The problem is most likely due to not pre-processing the test image properly (i.e. not using the same preprocessing pipeline as in training phase). That's why the model is producing such a sharp peak in the output values. — today

today today · Accepted Answer · 2019-04-20T19:55:07

The problem is that you are not applying the pre-processing pipeline you have used during the training of the model. Specifically, you must scale the values of the image pixels by 1/255.:

img = img.astype('float32') / 255.

It's really important to follow the same pre-processing pipeline in test phase as it is used the training phase; otherwise your model may get confused and outputs wrong predictions.

Why are the prediction probabilities so high for images with no defined class?

1 Answers