【deeplearning tips】dice loss解决分割正负样本不均衡

dice loss可以理解为检测任务重的IOU，用于分割任务中可以有效缓解正样本占比小的case。比如边缘的分割、小目标的分割。

前景分割：

#https://stackoverflow.com/questions/72195156/correct-implementation-of-dice-loss-in-tensorflow-keras
def dice_coef(y_true, y_pred, smooth=100):        
    y_true_f = K.flatten(y_true)
    y_pred_f = K.flatten(y_pred)
    intersection = K.sum(y_true_f * y_pred_f)
    dice = (2. * intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)
    return dice


def dice_coef_loss(y_true, y_pred):
    return 1-dice_coef(y_true, y_pred)

边缘+前景分割：

edge_kernel = np.array([1,1,1,1,-8,1,1,1,1], dtype=np.float32).reshape([3,3,1,1])
edge_kernel_const = tf.constant(edge_kernel)


def edge_dice_coef_loss(y_true, y_pred):
    y_true_edge = tf.nn.conv2d(y_true, edge_kernel_const, strides=1, padding='SAME')
    y_true_edge = 1. - tf.cast(tf.equal(y_true_edge, 0), dtype=tf.float32)#gt只有0、1
    
    y_pred_edge = tf.nn.conv2d(y_pred, edge_kernel_const, strides=1, padding='SAME')
    y_pred_edge = tf.pow(y_pred_edge, 2)#拉大边缘和非边缘区的值
    y_pred_edge = y_pred_edge/(y_pred_edge + 0.01) # 6/(6+0.01)=0.99, 0.01/(0.01+0.01)=0.5, 0.001/(0.001+0.01)=0.0    
    
    dice_loss = 1-dice_coef(y_true_edge, y_pred_edge)
    
    
    bce = tf.keras.losses.BinaryCrossentropy()
    bce_loss = bce(y_true, y_pred)
    return dice_loss + bce_loss