Spatially Variant Autoencoder for Unsupervised Pill Defect Detection

Abstract

In this thesis, we propose a spatially-variant autoencoder to differentiate between normal and defective medical pills with high accuracy, when it is hard to collect the images of defective samples. While most conventional methods rely on the existence of both normal and defective training data, the spatially variant autoencoder presented in this thesis is capable of distinguishing normal and defective pills, even though it is trained only on the normal samples. In the case of conventional autoencoders, every pixel is connected to the same pool of weights and biases in the fully-connected(FC) layer. Hence, in the case of a large input size, underfitting from the training data and test data will occur with a small number of parameters. If a relatively small number of weights are used for processing a large image with a high input-dimension, the output will be a low-quality image. Moreover, if the capacity of the network is too small, then the trained autoencoder sometimes outputs the same image all the time, regardless of the input image. On the other hand, if the number of parameters is increased to solve this problem, the output will become a near replica of the input image, even for the defective images that are not included as training samples. Unlike the conventional autoencoder, the proposed method in this thesis divides the fully connected inputs into very small patches, then carries out computations based on individual weight and bias for each patch. This process can also be seen as using multiple small autoencoders. In this way, the capacity of the network, as well as the computational complexity, can be minimized while the quality of the reconstructed images is not degraded. More importantly, this strongly limits the reconstruct ability of the autoencoder to the trained samples, which can be effective for distinguishing normal and defective samples. The difference between the input and output values of the spatially variant autoencoder can be calculated in order to distinguish between the normal and defective pills. In the experimental section, we will show the results for various cases and determine the appropriate size of the proposed autoencoder for the pills dataset.