Fix custom classifier false positives by removing BatchNormalization layers (#824)

Resolves train-inference mismatch that caused ~100% confidence scores on
background noise in custom classifiers trained with v2.0.0+.

Root cause: BatchNormalization layers were added in v2.0.0 but received
different input distributions during training (L2-normalized embeddings)
vs inference (raw embeddings), breaking learned decision boundaries.

Solution: Remove BatchNormalization layers from build_linear_classifier()
to restore reliable v1.5.1 behavior while keeping other v2.x improvements
(L2 regularization, improved learning rate schedule).

Fixes #823

Author: tphakala

birdnet_analyzer/model.py

@@ -635,9 +635,6 @@ def build_linear_classifier(num_labels, input_size, hidden_units=0, dropout=0.0)
     # Input layer
     model.add(keras.layers.InputLayer(input_shape=(input_size,)))
 
-    # Batch normalization on input to standardize embeddings
-    model.add(keras.layers.BatchNormalization())
-
     # Optional L2 regularization for all dense layers
     regularizer = keras.regularizers.l2(1e-5)
 
@@ -650,9 +647,6 @@ def build_linear_classifier(num_labels, input_size, hidden_units=0, dropout=0.0)
         # Add a hidden layer with L2 regularization
         model.add(keras.layers.Dense(hidden_units, activation="relu", kernel_regularizer=regularizer, kernel_initializer="he_normal"))
 
-        # Add another batch normalization after the hidden layer
-        model.add(keras.layers.BatchNormalization())
-
     # Dropout layer before output
     if dropout > 0:
         model.add(keras.layers.Dropout(dropout))