|
from keras.models import load_model |
|
import numpy as np |
|
|
|
labels = ['Chalky Soil', 'Mary Soil', 'Sand', 'Slit SOil', 'Alluvial Soil', 'Black Soil', 'Clay Soil', 'Red Soil'] |
|
MODEL_PATH = r"DenseNet121v2_95.h5" |
|
|
|
def load_CNN_model(): |
|
model = load_model(MODEL_PATH) |
|
print("model loaded") |
|
return model |
|
|
|
def classify_image(img): |
|
model = load_CNN_model() |
|
prediction = model.predict(img) |
|
predicted_class = np.argmax(prediction) |
|
|
|
predicted_label = labels[predicted_class] |
|
accuracy = prediction[0][predicted_class] |
|
|
|
return predicted_label, accuracy |
|
|