Alzheimer’s disease (AD) is a neurocognitive disorder that slowly impairs cognitive function and ultimately results in loss of individual function. Early prognosis improves the chances of preserving patient’s level of function and delays the onset of Dementia symptoms. Electroencephalogram (EEG) plays an important role in diagnosing patients with AD and offers a noninvasive continuous assessment of the disease progression. EEG signals collected from openly available dataset is used in this work. The dataset comprises four stages of Alzheimer: Normal Control (NC), Subjective Cognitive Decline (SCD), Mild Cognitive Impairment (MCI) and AD. Conversion of 1-D time-series to 2-D binary images without manual feature extraction is applied less on EEG data for AD diagnosis. Also, the nonlinear and non-stationary nature makes it a challenging task to interpret, examine and extract essential features from the EEG data for classification. The main objective of this study is to present a novel image transformation approach for an efficient AD diagnosis from EEG signals. For this, deep features are extracted from the generated images and are subsequently fed into Keras pre-trained deep learning models. While most of the existing studies performed binary classification of Alzheimer’s stages (NC vs AD), this work aims to perform multi-class classification on the EEG signal transformed into images. ResNet- 101v2 outperformed all other models with a highest accuracy of 98.98%. Thus, converting EEG signal into images and using the resulting images to train Convolutional Neural Networks (CNNs) for classification represents a very promising approach to AD diagnosis.