Early detection and diagnosis of arrhythmias, which results in cardiovascular diseases, is important for example with respect to Electrocardiogram (ECG) analysis. However, Convolutional Neural Networks (CNNs) along with other traditional deep learning models frequently fail for the processing of spatial hierarchies and viewpoint variations in ECG signals. In this research, we present an improved arrhythmia classification for which Capsule Networks (CapsNets) and scalogram-based feature extraction have been integrated together, where the time frequency representation of the ECG signals is used for better precision of classifying. The primary dataset, which is used, is MIT-BIH Arrhythmia Database, ECG signals are transformed into scalograms by Continuous Wavelet Transform (CWT) to be able to preserve the major morphological features. Based on this, CapsNet architecture is also developed to take advantage of spatial relationships and dynamic routing in CapsNet to address the limitations of CNNs in recognizing complex patterns. Experimental results show that a CNN-based approach obtains 98.75% average estimation accuracy for five arrhythmia classes (Normal (N), Supraventricular ectopic (S), Ventricular ectopic (V), Fusion beat (F), and Unknown (Q) beats) which beats 98.75%, which is also higher than 98.47% obtained by a conventional CNN-based approach. The proposed model is found to be superior in terms of the generalizability and robustness in comparison with the existing models. On the one hand, this research points to that CapsNets have potential applications in medical AI, and on the other hand, could enable real time, high precision, high interpretability and high reliability ECG monitoring systems.