L2 - Anomaly Detection in Computer Vision

channel like we discussed the last time considering this anime detection paper this keep gun anomaly we talked about the first component of the architecture i was using this paper which is the encoder now in this tutorial we'll be talking about the the second part of the architecture which is a decoder so let's look at what they formulated as the decoder okay so essentially what they did with the decoder network as you can see here is they took the bottleneck latent representation and they have sampled that back into the original reconstructed image which they termed the x-bar so in the bottom line they actually use 512 which is an hyper parameter as a latent vector so they did something creative here they borrowed idea from previous paper by using what you look at as a skip connection so that every uh level of this u-shaped decoder architecture they concatenated the features into the the up sample part of it which is the decoder part of it and the author claims that by using this keep skip connection actually improves the performance and skip connection it's also used in the popular resonant architecture and that has also proved to improve performance in the architecture so let's look at the loss function which they were trying to optimize so they used the loss function l1 loss and this x is the input image the x bar is a reconstruction and the x is sampled from the probability distribution of the normal samples so and that's the reconstruction loss the essence of this which they call the contextual laws is such that the the network will be able to actually learn the latent distribution of the input data for the normal samples because it is tax to reconstruct the number samples on the other hand of the decoder sample and let's see the architecture how they describe it quickly so they talked about the architecture [Music] here in this pipeline so that's what they refer to as the g g d g subscript d that's the decoder which like i said earlier up samples the latent vector z back into the image and we construct that at the target expat so they use the ib adopt the skip connection approach like i said and the skip connection helps to use a better reconstruction all right that's it for this session in the next session we continue to look at the other type of losses the adversary loss as well as the latent loss and we look at some more details with the training program thank you very much and see you in the next one