def LRF3DCNN(nClasses, H, W, Samples):
_input = Input(shape=(Samples, H, W, 16), name='LRF_Input')
lrf = Conv3D(32, (5, 2, 2), strides=(4, 2, 2), padding='same')(_input)
lrf = BatchNormalization(axis=-1)(lrf)
lrf = Activation('elu')(lrf)
# lrf = SpatialDropout3D(0.5)(lrf)
lrf = Conv3D(64, (5, 2, 2), strides=(4, 2, 2), padding='same')(lrf)
lrf = BatchNormalization(axis=-1)(lrf)
lrf = Activation('elu')(lrf)
# lrf = SpatialDropout3D(0.5)(lrf)
flatten = Flatten()(lrf)
dense = Dense(32)(flatten)
dense = BatchNormalization()(dense)
dense = Activation('relu')(dense)
dense = Dense(32)(dense)
dense = BatchNormalization()(dense)
dense = Activation('relu')(dense)
dense = Dense(nClasses)(dense)
_output = Activation('softmax', name='LRF_Output')(dense)
return Model(inputs=_input, outputs=_output, name='LRF')
def MB3DCNN(nClasses, H, W, Samples):
_input = Input(shape=(Samples, H, W, 1), name='MB_Input')
mb = Conv3D(16, (5, 3, 3), strides=(4, 2, 2), padding='same')(_input)
mb = BatchNormalization(axis=-1)(mb)
mb = Activation('elu')(mb)
# mb = SpatialDropout3D(0.5)(mb)
_srf_output = SRF3DCNN(nClasses, mb.shape[2], mb.shape[3], mb.shape[1])(mb)
_mrf_output = MRF3DCNN(nClasses, mb.shape[2], mb.shape[3], mb.shape[1])(mb)
_lrf_output = LRF3DCNN(nClasses, mb.shape[2], mb.shape[3], mb.shape[1])(mb)
_add = Add()([_srf_output, _mrf_output, _lrf_output])
# _output = Activation('softmax', name='MB_Output')(_add)
_output = Dense(nClasses, activation='softmax', name='MB_Output')(_add)
return Model(inputs=_input, outputs=[_output], name='MB3DCNN')