Anomaly detection on MNIST stroke simulation dataset

MNIST handwritten digit dataset is one of the most popular datasets in the filed of machine learning and pattern recognition, whereas several other datasets have also been synthesized using the MNIST dataset. Anomaly detection is an admired area of research in the field of computer vision. In this context, the present research offers an anomaly detection technique using the method of next frame prediction in MNIST hand stroke simulation dataset. The MNIST hand stroke simulation dataset is a synthesized dataset that tries to mimic the hand stroke sequences that humans mostly use while writing digits. After the generation of synthesized dataset, anomaly is introduced in this dataset as a Gaussian noise in randomly selected frames for every simulation. Later, the anomaly detection task is accomplished done using a three layer neural network architecture where the first layer comprises convolutional neural network (CNN or ConvNet) for appearance encoding and the second layer comprises ConvGRU for memorizing all past frames which correspond to the motion information. Finally, the third layer is a deconvolutional network, which is a decoder and tries to generate the next frame. The derived dataset used in this paper is created using the $K$-Means clustering algorithm, along with a handcrafted features to help in identification of the sequence in which the clusters are chosen during the simulation. Finally, the performance of the proposed model is compared with the state-of-the-art recurrent neural network architectures.