基于ResNet50的二分类目标检测CNN模型训练后异常输出极值边界框问题求助

各位大佬好！我最近在做一个二分类目标检测任务，用Keras的ResNet50做迁移学习搭建了CNN模型，但遇到了棘手的问题——不管我训练多少次（已经跑过好几轮100+epoch了），模型输出的边界框坐标总是固定的[0,0,0,1]，对应到我300x300的输入图就是Xmin、Ymin、Xmax全为0，Ymax为300，完全不符合真实标注。

先跟大家同步下我的数据情况：

• 输入图像形状是(batch size, 300, 300, 3)
• 边界框标签已经做了0-1归一化处理，形状为(batch size, 4)，打印标注数据能看到正常的结果，比如某一行是[0.74666667 0.32333333 0.88333333 0.69 ]
• 分类标签是独热编码格式，形状为(batch size, 2)

一开始我怀疑是DIoU损失函数的问题，换成MSE损失后，模型还是输出同样的极值边界框，看来不是损失函数的锅。下面是我的模型架构和DIoU损失函数的代码，麻烦各位帮忙排查下问题所在！

模型架构代码 #

from tensorflow.keras import layers
from tensorflow.keras import models
from keras.applications.resnet50 import ResNet50 


res = ResNet50(weights ='imagenet', include_top = False,  
               input_shape =(300, 300, 3))  
x = res.output  
x = layers.MaxPooling2D((2, 2))(x)

# Flatten and Fully connected layers
x = layers.BatchNormalization()(x)  

x = layers.Flatten()(x)
x = layers.Dropout(0.5)(x)
x = layers.Dense(512, activation='relu')(x)
x = layers.BatchNormalization()(x)  
x = layers.Dropout(0.50)(x)
x = layers.Dense(512)(x)

# Output layers with proper names
bbox_output = layers.Dense(4, activation='sigmoid', name='bbox_output')(x)  # Bounding box output
class_output = layers.Dense(2, activation='softmax', name='class_output')(x)  # Class output

# Define the model
model = models.Model(inputs = res.input, outputs=[bbox_output, class_output])

model.compile(
    optimizer='Adam',
    loss={'bbox_output': diou_loss, 'class_output': 'categorical_crossentropy'},
    metrics={'bbox_output': 'MSE', 'class_output': 'accuracy'})

callbacks = [
    keras.callbacks.EarlyStopping(
        # Stop training when `val_loss` is no longer improving
        monitor='val_bbox_output_loss',
        # "no longer improving" being defined as "no better than 1e-2 less"
        min_delta=1e-2,
        # "no longer improving" being further defined as "for at least 2 epochs"
        patience=20,
        verbose=1,
    )]

history = model.fit(X,
    [bbox_labels,class_labels], 
    epochs=100,batch_size=50, 
    verbose=1,
    validation_data=(X_test,[bbox_labels_test,class_labels_test]),
    callbacks=callbacks)

DIoU损失函数代码 #

def diou_loss(y_true, y_pred, epsilon=1e-7):
    # Use fixed image dimensions (256x256)

    
    print('yt',y_true)
    print('yp',y_pred)

    x_min_inter = tf.maximum(y_true[..., 0], y_pred[..., 0])
    y_min_inter = tf.maximum(y_true[..., 1], y_pred[..., 1])
    x_max_inter = tf.minimum(y_true[..., 2], y_pred[..., 2])
    y_max_inter = tf.minimum(y_true[..., 3], y_pred[..., 3])
    
    inter_area = tf.maximum(0.0, x_max_inter - x_min_inter) * tf.maximum(0.0, y_max_inter - y_min_inter)
    print('ia',inter_area)
    true_area = tf.maximum(0.0, y_true[..., 2] - y_true[..., 0]) * tf.maximum(0.0, y_true[..., 3] - y_true[..., 1])
    pred_area = tf.maximum(0.0, y_pred[..., 2] - y_pred[..., 0]) * tf.maximum(0.0, y_pred[..., 3] - y_pred[..., 1])
    union_area = true_area + pred_area - inter_area
    print('ua',union_area)
    iou = inter_area / tf.maximum(union_area, epsilon)
    print('iou',iou)

    # Calculate the center coordinates of the true and predicted boxes
    true_center_x = (y_true[..., 0] + y_true[..., 2]) / 2.0
    true_center_y = (y_true[..., 1] + y_true[..., 3]) / 2.0
    pred_center_x = (y_pred[..., 0] + y_pred[..., 2]) / 2.0
    pred_center_y = (y_pred[..., 1] + y_pred[..., 3]) / 2.0

    # Calculate the squared Euclidean distance between the centers
    center_distance = (true_center_x - pred_center_x) ** 2 + (true_center_y - pred_center_y) ** 2
    print('cd',center_distance)
    # Calculate the coordinates of the smallest enclosing box
    x_min_enclosing = tf.minimum(y_true[..., 0], y_pred[..., 0])
    y_min_enclosing = tf.minimum(y_true[..., 1], y_pred[..., 1])
    x_max_enclosing = tf.maximum(y_true[..., 2], y_pred[..., 2])
    y_max_enclosing = tf.maximum(y_true[..., 3], y_pred[..., 3])

    # Calculate the diagonal length squared of the enclosing box
    enclosing_diagonal = (x_max_enclosing - x_min_enclosing) ** 2 + (y_max_enclosing - y_min_enclosing) ** 2
    print('ed',enclosing_diagonal)
    # Calculate the DIoU
     
    # Return the DIoU loss
    return 1.0 - iou  + ((center_distance) / tf.maximum(enclosing_diagonal, 1e-7))

我已经卡了好几天了，实在找不到问题出在哪，恳请各位大佬给点思路，谢谢大家！

备注：内容来源于stack exchange，提问作者Caleb Joseph