I really need help with this, I am working on an anomaly detector project using Autoencoder, the first part of the project is to build an autoencoder trained only on non anomalous images so when it is tested on anomalous images (images never seen), it will reconstruct the non anomalous image and then compare between the input image (anomalous image) and the output image ( reconstructed anomaly-free image) so I can visualize the loss. but in my case the autoencoder is always reconstructing the input image as it is, whether it is anomaly free or anomalous image, in my project the autoencoder is not supposed to reconstruct the anomalous images as they are ( data never trained on ). Can someone please tell me what I am missing..

My training dataset is composed of 10 RGB identical anomaly-free images and my test dataset is composed of 1 anomalous image and 1 anomaly free image you can find below the output of the autoencoder for the test images.

https://i.stack.imgur.com/dj8cz.jpg

here's my autoencoder architecture I was aiming to modify it for better reconstruction but I stopped when it started reconstructing the anomalous images instead of anomaly-free images

def architecture_MVTEC(input_shape=(128, 128, 1), latent_dim=100):

parameters = dict()
N_layers = 8
parameters["filters"] = [32,  32, 64, 64, 128, 64, 32, latent_dim]
parameters["kernel_size"] = [4, 3, 4, 3, 4, 3, 3, 8]
parameters["strides"] = [2,1, 2, 1, 2, 1, 1, 1]
parameters["padding"] = ["same" for _ in range(N_layers-1)] + ["valid"]

# Input
inputs = Input(shape=input_shape)
x = inputs

# Encoder
for i in range(0, N_layers):
    x = Conv2D(
      filters=parameters["filters"][i],
      kernel_size=parameters["kernel_size"][i],
      strides=parameters["strides"][i],
      padding=parameters["padding"][i])(x)
    x = LeakyReLU(alpha=0.2)(x)

# Decoder
for i in reversed(range(0, N_layers)):
    x = Conv2DTranspose(
      filters=parameters["filters"][i],
      kernel_size=parameters["kernel_size"][i],
      strides=parameters["strides"][i],
      padding=parameters["padding"][i])(x)
    x = LeakyReLU(alpha=0.2)(x)

# Output
x = Conv2DTranspose(
  filters=input_shape[2],
  kernel_size=(3, 3),
  strides=(1, 1),
  padding="same")(x)

outputs = x

# Autoencoder
autoencoder = Model(inputs, outputs, name="autoencoder")[![enter image description here][1]][1]

return autoencoder

enter image description here

here's the autoencoder's summary :

Model: "autoencoder"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_1 (InputLayer)        [(None, 128, 128, 1)]     0         
                                                                 
 conv2d (Conv2D)             (None, 64, 64, 32)        544       
                                                                 
 leaky_re_lu (LeakyReLU)     (None, 64, 64, 32)        0         
                                                                 
 conv2d_1 (Conv2D)           (None, 32, 32, 32)        16416     
                                                                 
 leaky_re_lu_1 (LeakyReLU)   (None, 32, 32, 32)        0         
                                                                 
 conv2d_2 (Conv2D)           (None, 32, 32, 32)        9248      
                                                                 
 leaky_re_lu_2 (LeakyReLU)   (None, 32, 32, 32)        0         
                                                                 
 conv2d_3 (Conv2D)           (None, 16, 16, 64)        32832     
                                                                 
 leaky_re_lu_3 (LeakyReLU)   (None, 16, 16, 64)        0         
                                                                 
 conv2d_4 (Conv2D)           (None, 16, 16, 64)        36928     
                                                                 
 leaky_re_lu_4 (LeakyReLU)   (None, 16, 16, 64)        0         
                                                                 
 conv2d_5 (Conv2D)           (None, 8, 8, 128)         131200    
                                                                 
 leaky_re_lu_5 (LeakyReLU)   (None, 8, 8, 128)         0         
                                                                 
 conv2d_6 (Conv2D)           (None, 8, 8, 64)          73792     
                                                                 
 leaky_re_lu_6 (LeakyReLU)   (None, 8, 8, 64)          0         
                                                                 
 conv2d_7 (Conv2D)           (None, 8, 8, 32)          18464     
                                                                 
 leaky_re_lu_7 (LeakyReLU)   (None, 8, 8, 32)          0         
                                                                 
 conv2d_8 (Conv2D)           (None, 1, 1, 100)         204900    
                                                                 
 leaky_re_lu_8 (LeakyReLU)   (None, 1, 1, 100)         0         
                                                                 
 conv2d_transpose (Conv2DTra  (None, 8, 8, 100)        640100    
 nspose)                                                         
                                                                 
 leaky_re_lu_9 (LeakyReLU)   (None, 8, 8, 100)         0         
                                                                 
 conv2d_transpose_1 (Conv2DT  (None, 8, 8, 32)         28832     
 ranspose)                                                       
                                                                 
 leaky_re_lu_10 (LeakyReLU)  (None, 8, 8, 32)          0         
                                                                 
 conv2d_transpose_2 (Conv2DT  (None, 8, 8, 64)         18496     
 ranspose)                                                       
                                                                 
 leaky_re_lu_11 (LeakyReLU)  (None, 8, 8, 64)          0         
                                                                 
 conv2d_transpose_3 (Conv2DT  (None, 16, 16, 128)      131200    
 ranspose)                                                       
                                                                 
 leaky_re_lu_12 (LeakyReLU)  (None, 16, 16, 128)       0         
                                                                 
 conv2d_transpose_4 (Conv2DT  (None, 16, 16, 64)       73792     
 ranspose)                                                       
                                                                 
 leaky_re_lu_13 (LeakyReLU)  (None, 16, 16, 64)        0         
                                                                 
 conv2d_transpose_5 (Conv2DT  (None, 32, 32, 64)       65600     
 ranspose)                                                       
                                                                 
 leaky_re_lu_14 (LeakyReLU)  (None, 32, 32, 64)        0         
                                                                 
 conv2d_transpose_6 (Conv2DT  (None, 32, 32, 32)       18464     
 ranspose)                                                       
                                                                 
 leaky_re_lu_15 (LeakyReLU)  (None, 32, 32, 32)        0         
                                                                 
 conv2d_transpose_7 (Conv2DT  (None, 64, 64, 32)       16416     
 ranspose)                                                       
                                                                 
 leaky_re_lu_16 (LeakyReLU)  (None, 64, 64, 32)        0         
                                                                 
 conv2d_transpose_8 (Conv2DT  (None, 128, 128, 32)     16416     
 ranspose)                                                       
                                                                 
 leaky_re_lu_17 (LeakyReLU)  (None, 128, 128, 32)      0         
                                                                 
 conv2d_transpose_9 (Conv2DT  (None, 128, 128, 1)      289       
 ranspose)                                                       
                                                                 
=================================================================
Total params: 1,533,929
Trainable params: 1,533,929
Non-trainable params: 0
_________________________________________________________________