Keras-TF model.predict() 给了我错误的结果

Question

I just trained my CNN model which has the layer details below:

model = Sequential()
model.add(Conv2D(100, (3, 3), input_shape=(100, 100, 1), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(120, (5, 5), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(140, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(rate=0.5))
model.add(Flatten())
model.add(Dense(200, activation='relu'))
model.add(Dropout(rate=0.5))
model.add(Dense(100, activation='relu'))
model.add(Dropout(rate=0.5))
model.add(Dense(num_classes, activation='softmax'))
model.compile(Adam(lr=0.001), loss='categorical_crossentropy', metrics=['accuracy'])

My model has about 95.xxx% accuracy on keras validating/testing result.

And when I imported the h5 file to test the prediction of an image, I did:

def grayscale(img):
  img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
  return img

def equalize(img):
  img = cv2.equalizeHist(img)
  return img

def preprocessing(img):
  gray = grayscale(img)
  eq = equalize(gray)
  return eq / 255

image = cv2.imread("sample.jpg")
print("Loading model...")
model = load_model("model.h5")
classes = pd.read_csv('dataset.csv', header=0, usecols=['BananaOrNot']).values

image = preprocessing(image)
image = image.reshape(100, 100, 1)
image = np.expand_dims(image, axis=0)
y_prob = model.predict(image)
class_idx = y_prob.argmax(axis=-1)[0]
print(classes[class_idx][0]) # it produced wrong result (should be 'not banana', got 'banana')

I tested with a sample blank white image, and the prediction gave me Banana result, even though it should be Not banana for sure. How could it be wrong in this very simple test case? Or is there anything wrong with my model.predict(img) input image?

Answer 1

Since there is no class named Not banana in your dataset, it gives only the classes that you've trained in your network.

In order to have that behavior on your model, you should set a class e.g: Background on your model and train it with certain amount of Background labeled images. Otherwise it would give the result that most probable one in your output vector i.e the most probable class even if it's a blank image.

OR

You can train a classification model along with the localization and classification only occurs in localized objects and your problem may be corrected.