HP INCFollow


Deep Learning (DL) applications usually need a large amount of labeled data to perform successfully. Data augmentation is

a well-known method to avoid overfitting during a training process by increasing the number of training images with simple

transformations that do not modify the true nature of the sample. However, these modifications are usually handcrafted and

problem-dependent as it is difficult to determine which one is better. For these reasons, there is an increasing interest in

studying automatic mechanisms to generate new instances using the available data. Variational Autoencoders (VAE) are

powerful generative models. They’re a modern take on autoencoders—a type of network that aims to encode an input to a lowdimensional

latent space and then decode it back—that mixes ideas from deep learning with Bayesian inference. VAEs are

usually easier to train compared to Generative Adversarial Nets (GANs), although the samples are sometimes blurrier.

In this work, we present VAE based print defect generator that requires minimum (few dozen, e.g. 20-30) samples per defect

type. First, we design a small compact Fully Convolutional Network (FCN) inspired by "Small U-Net" which is encoderdecoder

type network used for detecting vehicles in a video stream [

master/]. This is mainly achieved by using Separable convolutions which separates the learning of spatial

features and channel-wise features and using global max pooling instead of fully-connected layers. In order to improve further

the generalization capabilities of our network (which is called CompactSmallU-Net) we use simple data augmentation (flipping,

color channels swap, etc.), drop-out and batch normalization while training the VAE. Next, we create a database (DB) of

images with real looking defects by using seamless cloning [] to overlay

the generated defects on top of (defects free) images. Finally, Siamese based Single Shot Detector (SSD) can be used to train

a press inspection system (AAA- Automatic Alert Agent) for detecting and classifying print defects (e.g., drips & dents) given

pair of RGB images (digital and scanned) from the DB.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Share Alike 4.0 License.