Bio-classification and analysis tool design and implementation

The problem of recognizing/classifying shapes or images is a subject that has been studied for many years. The analysis of images obtained in different disciplines such as biological species, tomography, x-ray images may require researchers working in different disciplines to have technical knowledge about image processing methods that are not parts of their own discipline. Within the scope of this study, an easy-to-use, convolutional neural network-based feature extraction and processing toolset that can be used in different disciplines is designed and implemented. The library and the user interface are tested on popular data sets related to different disciplines. Tests on different data sets using pre-trained convolutional neural networks show that the tools developed to extract and process features obtained using convolutional neural networks with the concept of transfer learning can successfully extract distinctive features. Various technologies of processors such as multi-core, "single instruction, multiple data" are also used in order for the functions to run as fast as possible, in the implementations of the presented toolset. The processing speed is increased drastically by optimizing codes and making parallelization on the processor cores. Solutions are also presented for problems such as shared resource utilization, encountered during implementation. Using the toolset, relations between layer depths and distinctivenes of features in deep convolutional neural networks are also investigated and it is shown that features obtained from different layers when reduced to sub-feature sets based on correlations between features and classes, can provide better results than the outputs of the entire network.