Flexible Image Recognition Software Toolbox (First)

Abstract

The deep convolutional neural network is a new concept in the neural network field, and research is still going on to improve network performance. These networks are used for recognizing patterns in data, as they provide shift invariance, automatic extraction of local features, by using local receptive fields, and improved generalization, by using weight sharing techniques. The main purpose of this thesis is to create a Flexible Image Recognition Software Toolbox (FIRST), which is a software package that allows users to build custom deep networks, while also having ready-made versions of popular deep networks, such as Lecun's LeNet-1, LeNet-5 and LeNet-7. This software package is created for designing, training and simulating deep networks. The goal is to reduce the amount of time required by users to implement any particular network. To design this software package, a general modular framework is introduced, in which simulation and gradient calculations are derived. Due to this modularity and generality, FIRST provides flexibility to users in easily designing specific complex or deep networks. FIRST includes several training algorithms, such as Resilient Backpropagation, Scaled Conjugate Gradient and Steepest Descent. This thesis also describes the usage of the FIRST software and the design of functions used in the software. It also provides information about how to create custom networks. The thesis includes two sample training sessions that demonstrate how to use the FIRST software. One example is phoneme recognition in 1D speech data. The second example is handwritten digit recognition in 2D images.