Obtaining information signal from modulation signal using machine learning method and image processing

Machine Learning and Image Processing algorithms are used in various ways in the classification or recognition of communication signals. Currently, thanks to the diversity of digital communication signals, faster communication is achieved in a lower bandwidth. The ability to separate or demodulate the signals, also called passband modulation, used in digital communication, to detect the signal easily and quickly at the receiver side and to generate response accordingly are also important in both military and civilian applications. Within the scope of the thesis, instead of demodulator circuit structures, images of modulated QASK, QFSK and QPSK the classification and demodulation processes of these signals are performed by taking the images of the signals. In order to classify and demodulate signals, support vector machines (SVM), convolutional neural networks (CNN), combined model (Ensemble), NARX artificial neural network structure and Oriented FAST and Rotated BRIEF (ORB) method were used to extract features from images. As a result of classification and demodulation of images consisting of modulated signals with 0dB, 5dB, 10dB and 15dB noise ratios, their performances are evaluated and presented with statistical parameters. It showed the confusion matrices included in the results and complexities occurred in the prediction of signal images, especially with 5dB noise ratio. It has been shown that this situation can be overcome by getting support from feature extraction algorithms or by combining machine learning methods.