Abstarct: The purpose of current research is to study font recognition in order to identify font from the scanned images of documents. Font recognition is an important part of OCR (Optical Character Recognition) system which aims to recognize the font used in a typed text. Font recognition has been implemented in this thesis by combining different characteristics and using SVM (Support Vector Machines) classifier and MLP (Multi Linear Perception) network. This study utilizes Gabor filter, wavelet transform, multi-dimensional fractal and SRF (Sobel-Roberts Feature) which is a gradient-baxsed method to extract features on the images of fonts and then compares them with each other. SRF method is a combination of Sobel and Roberts masks which employs amplitude and phase data for feature extraction. Meanwhile, features are extracted in various directions by this technique. Moreover, different mother wavelets are used and compared for the purpose of feature extraction, including Dbi and Symlet. Another feature extraction method is baxsed on fractal. Fractal is a useful method to determine complexities of the images. Several techniques such as box counting and Dilation counting are utilized and assessed to calculate dimension of the fractal in this research. Gabor filter is another way of extracting features which has been used extensively before. One major disadvantage of Gabor filter is related to its heavy and time-consuming calculations, whereas other methods involve much fewer calculations and perform more quickly than Gabor filter for feature extraction. Having extracted the features from different methods, they are recognized by SVM and MLP classifiers and their results are compared with each other. SRF and wavelet techniques show acceptable speeds and provide better results. On the other hand, these two features have numerous differences in how they extract features, thus their errors will demonstrate a rather small correlation. Taking into account this finding, these two features are combined in this work to yield significantly improved results. Average recognition rate for RSW (Roberts-Sobel-Wavelet) method using SVM classifier is found 95.58% which is 2.14% and 11.25% greater than those of SRF and wavelet transform methods, respectively.