基于空洞卷积的多尺度2D特征融合网络
Posted jinyiyexingzc
tags:
篇首语：本文由小常识网(cha138.com)小编为大家整理，主要介绍了基于空洞卷积的多尺度2D特征融合网络相关的知识，希望对你有一定的参考价值。
空洞卷积模块2，来替代编码网络中每一层的卷积，图3.4展示了空洞卷积模块结构，该模块共使用了五个不同尺度的卷积组，分别用扩张率1，2，4，8，16的空洞卷积进行组合，第一组为扩张率为1的3×3卷积，可以得到每一像素点感受野为3×3的特征图；第二组级联扩张率为1和2的3×3卷积，感受野为7×7；依此类推，后面每一组额外多并行一个2倍扩张率的空洞卷积，第三组、第四组、第五组卷积得到的感受野分别为15×15、31×31、63×63。经过修改的空洞卷积模块2的每一个并行的特征处理通道中，空洞卷积的参数覆盖率是百分百，不存在信息丢失的问题。在模块末端，将这五组不同尺度的特征图进行拼接，将拼接的特征图再通过一个普通卷积，上述操作实现了多尺度特征融合。
网络训练程序
1 import keras 2 from keras.models import * 3 from keras.layers import Input, Conv2D, MaxPooling2D, UpSampling2D, Dropout 4 from keras.optimizers import * 5 6 from keras.layers import Concatenate 7 8 from keras import backend as K 9 10 import matplotlib.pyplot as plt 11 from keras.callbacks import ModelCheckpoint 12 from fit_generator import get_path_list, get_train_batch 13 14 15 train_batch_size = 1 16 epoch = 1 17 18 data_train_path = "../deform/train_label_dir/train3" 19 data_label_path = "../deform/train_label_dir/label3" 20 21 train_path_list, label_path_list, count = get_path_list(data_train_path, data_label_path) 22 23 24 # 写一个LossHistory类，保存loss和acc 25 class LossHistory(keras.callbacks.Callback): 26 def on_train_begin(self, logs={}): 27 self.losses = {‘batch‘: [], ‘epoch‘:[]} 28 self.accuracy = {‘batch‘: [], ‘epoch‘:[]} 29 self.val_loss = {‘batch‘: [], ‘epoch‘:[]} 30 self.val_acc = {‘batch‘: [], ‘epoch‘:[]} 31 32 def on_batch_end(self, batch, logs={}): 33 self.losses[‘batch‘].append(logs.get(‘loss‘)) 34 self.accuracy[‘batch‘].append(logs.get(‘dice_coef‘)) 35 self.val_loss[‘batch‘].append(logs.get(‘val_loss‘)) 36 self.val_acc[‘batch‘].append(logs.get(‘val_acc‘)) 37 38 def on_epoch_end(self, batch, logs={}): 39 self.losses[‘epoch‘].append(logs.get(‘loss‘)) 40 self.accuracy[‘epoch‘].append(logs.get(‘dice_coef‘)) 41 self.val_loss[‘epoch‘].append(logs.get(‘val_loss‘)) 42 self.val_acc[‘epoch‘].append(logs.get(‘val_acc‘)) 43 44 def loss_plot(self, loss_type): 45 iters = range(len(self.losses[loss_type])) 46 plt.figure(1) 47 # acc 48 plt.plot(iters, self.accuracy[loss_type], ‘r‘, label=‘train dice‘) 49 if loss_type == ‘epoch‘: 50 # val_acc 51 plt.plot(iters, self.val_acc[loss_type], ‘b‘, label=‘val acc‘) 52 plt.grid(True) 53 plt.xlabel(loss_type) 54 plt.ylabel(‘dice‘) 55 plt.legend(loc="best") 56 plt.savefig(‘./curve_figure/tune_liver/unet_liver2_raw_0_129_dialtion_all_entropy_dice_curve2.png‘) 57 58 plt.figure(2) 59 # loss 60 plt.plot(iters, self.losses[loss_type], ‘g‘, label=‘train loss‘) 61 if loss_type == ‘epoch‘: 62 # val_loss 63 plt.plot(iters, self.val_loss[loss_type], ‘k‘, label=‘val loss‘) 64 plt.grid(True) 65 plt.xlabel(loss_type) 66 plt.ylabel(‘loss‘) 67 plt.legend(loc="best") 68 plt.savefig(‘./curve_figure/tune_liver/unet_liver2_raw_0_129_dialtion_all_entropy_loss_curve2.png‘) 69 plt.show() 70 71 72 def dice_coef(y_true, y_pred): 73 smooth = 1. 74 y_true_f = K.flatten(y_true) 75 y_pred_f = K.flatten(y_pred) 76 intersection = K.sum(y_true_f * y_pred_f) 77 return (2. * intersection + smooth) / (K.sum(y_true_f * y_true_f) + K.sum(y_pred_f * y_pred_f) + smooth) 78 79 80 def dice_coef_loss(y_true, y_pred): 81 return 1. - dice_coef(y_true, y_pred) 82 83 84 def mycrossentropy(y_true, y_pred, e=0.1): 85 nb_classes = 10 86 loss1 = K.categorical_crossentropy(y_true, y_pred) 87 loss2 = K.categorical_crossentropy(K.ones_like(y_pred) / nb_classes, y_pred) 88 return (1 - e) * loss1 + e * loss2 89 90 91 class myUnet(object): 92 def __init__(self, img_rows=512, img_cols=512): 93 self.img_rows = img_rows 94 self.img_cols = img_cols 95 96 def dilation_conv(self, kernel_num, kernel_size, input): 97 # 空洞卷积多尺度模块 98 conv1_1 = Conv2D(kernel_num, kernel_size, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘, 99 dilation_rate=(1, 1))(input) 100 conv1_2 = Conv2D(kernel_num, kernel_size, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘, 101 dilation_rate=(2, 2))(input) 102 conv1_3 = Conv2D(kernel_num, kernel_size, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘, 103 dilation_rate=(4, 4))(input) 104 conv1_4 = Conv2D(kernel_num, kernel_size, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘, 105 dilation_rate=(8, 8))(input) 106 conv1_5 = Conv2D(kernel_num, kernel_size, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘, 107 dilation_rate=(16, 16))(input) 108 merges = Concatenate(axis=3)([conv1_1, conv1_2, conv1_3, conv1_4, conv1_5]) 109 return merges 110 111 def BN_operation(self, input): 112 output = keras.layers.normalization.BatchNormalization(axis=-1, momentum=0.99, epsilon=0.001, center=True, 113 scale=True, 114 beta_initializer=‘zeros‘, gamma_initializer=‘ones‘, 115 moving_mean_initializer=‘zeros‘, 116 moving_variance_initializer=‘ones‘, 117 beta_regularizer=None, 118 gamma_regularizer=None, beta_constraint=None, 119 gamma_constraint=None)(input) 120 return output 121 122 def get_unet(self): 123 inputs = Input((self.img_rows, self.img_cols, 1)) 124 125 conv1 = self.dilation_conv(64, 3, inputs) 126 conv1 = Conv2D(64, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(conv1) 127 pool1 = MaxPooling2D(pool_size=(2, 2))(conv1) 128 # BN 129 pool1 = self.BN_operation(pool1) 130 131 conv2 = self.dilation_conv(128, 3, pool1) 132 conv2 = Conv2D(128, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(conv2) 133 pool2 = MaxPooling2D(pool_size=(2, 2))(conv2) 134 # BN 135 pool2 = self.BN_operation(pool2) 136 137 conv3 = self.dilation_conv(256, 3, pool2) 138 conv3 = Conv2D(256, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(conv3) 139 pool3 = MaxPooling2D(pool_size=(2, 2))(conv3) 140 # BN 141 pool3 = self.BN_operation(pool3) 142 143 conv4 = self.dilation_conv(512, 3, pool3) 144 conv4 = Conv2D(512, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(conv4) 145 drop4 = Dropout(0.5)(conv4) 146 pool4 = MaxPooling2D(pool_size=(2, 2))(drop4) 147 # BN 148 pool4 = self.BN_operation(pool4) 149 150 conv5 = Conv2D(1024, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(pool4) 151 conv5 = Conv2D(1024, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(conv5) 152 drop5 = Dropout(0.5)(conv5) 153 # BN 154 drop5 = self.BN_operation(drop5) 155 156 up6 = Conv2D(512, 2, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)( 157 UpSampling2D(size=(2, 2))(drop5)) 158 merge6 = Concatenate(axis=3)([drop4, up6]) 159 conv6 = Conv2D(512, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(merge6) 160 conv6 = Conv2D(512, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(conv6) 161 162 up7 = Conv2D(256, 2, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)( 163 UpSampling2D(size=(2, 2))(conv6)) 164 merge7 = Concatenate(axis=3)([conv3, up7]) 165 conv7 = Conv2D(256, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(merge7) 166 conv7 = Conv2D(256, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(conv7) 167 168 up8 = Conv2D(128, 2, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)( 169 UpSampling2D(size=(2, 2))(conv7)) 170 merge8 = Concatenate(axis=3)([conv2, up8]) 171 conv8 = Conv2D(128, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(merge8) 172 conv8 = Conv2D(128, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(conv8) 173 174 up9 = Conv2D(64, 2, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)( 175 UpSampling2D(size=(2, 2))(conv8)) 176 merge9 = Concatenate(axis=3)([conv1, up9]) 177 conv9 = Conv2D(64, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(merge9) 178 conv9 = Conv2D(64, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(conv9) 179 conv9 = Conv2D(2, 3, activation=‘relu‘, padding=‘same‘, kernel_initializer=‘he_normal‘)(conv9) 180 conv10 = Conv2D(1, 1, activation=‘sigmoid‘)(conv9) 181 182 model = Model(inputs=inputs, outputs=conv10) 183 184 # 在这里可以自定义损失函数loss和准确率函数accuracy 185 # model.compile(optimizer=Adam(lr=1e-4), loss=‘binary_crossentropy‘, metrics=[‘accuracy‘]) 186 model.compile(optimizer=Adam(lr=1e-4), loss= ‘binary_crossentropy‘, metrics=[‘accuracy‘, dice_coef]) 187 print(‘model compile‘) 188 return model 189 190 def train(self): 191 # model = self.get_unet() 192 model = load_model(‘./model/dilation_tune/unet_liver2_dir1_dilation_all_entropy.hdf5‘, custom_objects={‘dice_coef‘: dice_coef}) 193 194 195 print("got unet") 196 197 # 保存的是模型和权重 198 model_checkpoint = ModelCheckpoint(‘./model/dilation_tune/unet_liver2_dir2_dilation_all_entropy.hdf5‘, monitor=‘loss‘, 199 verbose=1, save_best_only=True) 200 print(‘Fitting model...‘) 201 202 # 创建一个实例history 203 history = LossHistory() 204 model.fit_generator( 205 generator=get_train_batch(train_path_list, label_path_list, train_batch_size, 512, 512), 206 epochs=epoch, verbose=1, 207 steps_per_epoch=count//train_batch_size, 208 callbacks=[model_checkpoint, history], 209 workers=1) 210 211 # 绘制acc-loss曲线 212 history.loss_plot(‘batch‘) 213 214 215 if __name__ == ‘__main__‘: 216 myunet = myUnet() 217 myunet.train()
模型测试、模型评估，同U-Net网络
以上是关于基于空洞卷积的多尺度2D特征融合网络的主要内容，如果未能解决你的问题，请参考以下文章