深度可分离卷积检测算法

Language: Python

使用TensorFlow 深度学习框架,使用Keras会大幅缩减代码量

训练机器:华为Atlas 200 AI开发板(或本地计算机)

数据集

常用的卷积网络模型及在ImageNet上的准确率

模型 大小 Top-1准确率 Top-5准确率 参数数量 深度
Xception 88 MB 0.790 0.945 22,910,480 126
VGG16 528 MB 0.713 0.901 138,357,544 23
VGG19 549 MB 0.713 0.900 143,667,240 26
ResNet50 98 MB 0.749 0.921 25,636,712 168
ResNet101 171 MB 0.764 0.928 44,707,176 -
ResNet152 232 MB 0.766 0.931 60,419,944 -
ResNet50V2 98 MB 0.760 0.930 25,613,800 -
ResNet101V2 171 MB 0.772 0.938 44,675,560 -
ResNet152V2 232 MB 0.780 0.942 60,380,648 -
ResNeXt50 96 MB 0.777 0.938 25,097,128 -
ResNeXt101 170 MB 0.787 0.943 44,315,560 -
InceptionV3 92 MB 0.779 0.937 23,851,784 159
InceptionResNetV2 215 MB 0.803 0.953 55,873,736 572
MobileNet 16 MB 0.704 0.895 4,253,864 88
MobileNetV2 14 MB 0.713 0.901 3,538,984 88
DenseNet121 33 MB 0.750 0.923 8,062,504 121
DenseNet169 57 MB 0.762 0.932 14,307,880 169
DenseNet201 80 MB 0.773 0.936 20,242,984 201
NASNetMobile 23 MB 0.744 0.919 5,326,716 -
NASNetLarge 343 MB 0.825 0.960 88,949,818 -

由于硬件条件限制,综合考虑模型的准确率、大小以及复杂度等因素,采用了Xception模型, 该模型是134层(包含激活层,批标准化层等)拓扑深度的卷积网络模型。

检测算法

def Xception(include_top=True,
    weights='imagenet',
    input_tensor=None,
    input_shape=None,
    pooling=None,
    classes=1000,
    **kwargs)

# 参数
# include_top:是否保留顶层的全连接网络
# weights:None代表随机初始化,即不加载预训练权重。'imagenet’代表加载预训练权重
# input_tensor:可填入Keras tensor作为模型的图像输入tensor
# input_shape:可选,仅当include_top=False有效,应为长为3的tuple,指明输入图片的shape,图片的宽高必须大于71,如(150,150,3)
# pooling:当include_top=False时,该参数指定了池化方式。None代表不池化,最后一个卷积层的输出为4D张量。‘avg’代表全局平均池化,‘max’代表全局最大值池化。
# classes:可选,图片分类的类别数,仅当include_top=True并且不加载预训练权重时可用

基于Xception的模型微调,详细请参考代码

  1. 设置Xception参数

    迁移学习参数权重加载:xception_weights

     # 设置输入图像的宽高以及通道数
     img_size = (299, 299, 3)
    
     base_model = keras.applications.xception.Xception(include_top=False,
                                                     weights='..\\resources\\keras-model\\xception_weights_tf_dim_ordering_tf_kernels_notop.h5',
                                                     input_shape=img_size,
                                                     pooling='avg')
    
     # 全连接层,使用softmax激活函数计算概率值,分类大小是628
     model = keras.layers.Dense(628, activation='softmax', name='predictions')(base_model.output)
     model = keras.Model(base_model.input, model)
    
     # 锁定卷积层
     for layer in base_model.layers:
       layer.trainable = False
    
  2. 全连接层训练(v1.0)

     from base_model import model
    
     # 设置训练集图片大小以及目录参数
     img_size = (299, 299)
     dataset_dir = '..\\dataset\\dataset'
     img_save_to_dir = 'resources\\image-traing\\'
     log_dir = 'resources\\train-log'
    
     model_dir = 'resources\\keras-model\\'
    
     # 使用数据增强
     train_datagen = keras.preprocessing.image.ImageDataGenerator(
         rescale=1. / 255,
         shear_range=0.2,
         width_shift_range=0.4,
         height_shift_range=0.4,
         rotation_range=90,
         zoom_range=0.7,
         horizontal_flip=True,
         vertical_flip=True,
         preprocessing_function=keras.applications.xception.preprocess_input)
    
     test_datagen = keras.preprocessing.image.ImageDataGenerator(
         preprocessing_function=keras.applications.xception.preprocess_input)
    
     train_generator = train_datagen.flow_from_directory(
         dataset_dir,
         save_to_dir=img_save_to_dir,
         target_size=img_size,
         class_mode='categorical')
    
     validation_generator = test_datagen.flow_from_directory(
         dataset_dir,
         save_to_dir=img_save_to_dir,
         target_size=img_size,
         class_mode='categorical')
    
     # 早停法以及动态学习率设置
     early_stop = EarlyStopping(monitor='val_loss', patience=13)
     reduce_lr = ReduceLROnPlateau(monitor='val_loss', patience=7, mode='auto', factor=0.2)
     tensorboard = keras.callbacks.tensorboard_v2.TensorBoard(log_dir=log_dir)
    
     for layer in model.layers:
         layer.trainable = False
    
     # 模型编译
     model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])
    
     history = model.fit_generator(train_generator,
                                   steps_per_epoch=train_generator.samples // train_generator.batch_size,
                                   epochs=100,
                                   validation_data=validation_generator,
                                   validation_steps=validation_generator.samples // validation_generator.batch_size,
                                   callbacks=[early_stop, reduce_lr, tensorboard])
     # 模型导出
     model.save(model_dir + 'chinese_medicine_model_v1.0.h5')
    
  3. 对于顶部的6层卷积层,我们使用数据集对权重参数进行微调

     # 加载模型
     model=keras.models.load_model('resources\\keras-model\\chinese_medicine_model_v2.0.h5')
    
     for layer in model.layers:
        layer.trainable = False
     for layer in model.layers[126:132]:
        layer.trainable = True
    
     history = model.fit_generator(train_generator,
                                   steps_per_epoch=train_generator.samples // train_generator.batch_size,
                                   epochs=100,
                                   validation_data=validation_generator,
                                   validation_steps=validation_generator.samples // validation_generator.batch_size,
                                   callbacks=[early_stop, reduce_lr, tensorboard])
     model.save(model_dir + 'chinese_medicine_model_v2.0.h5')
    
  4. 在后端项目中,我们使用Deeplearn4j调用训练好的模型 ``` public class CnnModelUtil {

     private static ComputationGraph CNN_MODEL = null;
    
     /**
      * 中药名字的编码
      */
     private static final Map<Integer, String> MEDICINE_NAME_MAP = new HashMap<>();
    
     /**
      * 定义cnn model的文件夹路径
      */
     private static final String DATA_DIR = System.getProperty("os.name")
             .toLowerCase().contains("windows") ? "D:\\data\\model\\"
             : "./data/model/";
    
     /**
      * 定义中药编码表的文件名
      */
     private static final String MEDICINE_LABLE_FILE_NAME = "medicine_name-lable.txt";
    
     /**
      * 定义模型的文件名
      */
     private static final String CNN_MODEL_FILE_NAME = "chinese_medicine_model.h5";
    
    /**
     * 图片的加载器
     */
    private static final NativeImageLoader IMAGE_LOADER = new NativeImageLoader(299, 299, 3);


    /**
     * 初始化
     */
    static {
        try {
            CNN_MODEL = KerasModelImport.importKerasModelAndWeights(DATA_DIR + CNN_MODEL_FILE_NAME);

            Files.readAllLines(Paths.get(DATA_DIR, MEDICINE_LABLE_FILE_NAME)).forEach(v -> {
                String[] split = v.split(",");
                MEDICINE_NAME_MAP.put(Integer.valueOf(split[1]), split[0]);
            });
        } catch (IOException | InvalidKerasConfigurationException | UnsupportedKerasConfigurationException e) {
            e.printStackTrace();
        }
    }

    /**
     * 对图像进行预测
     * 对预测的概率值进行排序处理
     * 返回值是概率值前10的中药的名字
     * @param file
     * @return
     * @throws 
     */
    public static Map<String, Float> medicineNamePredict(File file) throws IOException {
        INDArray image = IMAGE_LOADER.asMatrix(file).divi(127.5).subi(1);
        INDArray output = CNN_MODEL.outputSingle(image);
        Map<Integer, Float> resultMap = new HashMap<>();
        float[] floats = output.toFloatVector();
        for (int i = 0; i < floats.length; i++) {
            resultMap.put(i, floats[i]);
        }
        List<Map.Entry<Integer, Float>> resultList = new LinkedList<>(resultMap.entrySet());
        resultList.sort(Map.Entry.comparingByValue(Comparator.reverseOrder()));
        Map<String, Float> medicinePredict = new LinkedHashMap<>();
        resultList.stream().limit(10).forEach(v -> {
            medicinePredict.put(MEDICINE_NAME_MAP.get(v.getKey()), v.getValue());
        });
        return medicinePredict;
    }
}
```

模型概览

模型详细结构

训练过程正确率以及损失函数可视化展示

正确率 损失函数

results matching ""

    No results matching ""