The functional Magnetic Resonance Imaging (fMRI) is one among the non-invasive techniques used in cognitive neuroscience, to record the activity of the brain. The fMRI reveals the functional activity caused by the Blood Oxygenated Level Dependent (BOLD) signals in the brain. The visual image reconstruction allows to translate neural brain activity pattern into the visual image (stimulus) that has caused the corresponding brain activity. The stimulus may be graphical characters, face images, handwritten characters and natural images. The proposed technique aims to develop a novel framework for visual image reconstruction of natural images from fMRI. The exact reconstruction of natural images is challenging due to its complexity. The classification of image prior plays an important role in improving the accuracy of reconstruction. Since the image prior consists of multiple categories, like Gabor wavelet transform, Scale invariant feature transform (SIFT), Speeded Up Robust Features (SURF), Local Binary Pattern (LBP), Haar feature transform, Bag of Visual Words (BoVW) were tried on image prior. The extracted features are fed into the multiclass Support Vector Machine classifier followed by k-means clustering. An analysis on reconstruction done using different feature extraction techniques revealed that the Gabor feature extraction gave the highest accuracy in final results. The reconstruction of natural images was achieved with an accuracy of 80% till now. Also 70% accuracy was achieved in identifying the category and reconstructing the test stimulus from a real time test fMRI voxel responses. The proposed work focuses on developing an accurate, less complex and automatic software technique for visual image reconstruction of natural images.