EEG-based emotion recognition is an important task in affective computing; however, conventional machine learning methods are limited in reliably extracting emotional features because brain signals are highly non-stationary and noisy. To overcome the limitation, we present a deep learning framework that integrates CWT-based RGB scalogram generation and transfer learning with MobileNetV2 for reliable emotion classification using EEG. In this way, the raw EEG data is first standardized and transformed into time–frequency scalograms using the Morlet wavelet with different scales, which are then resized and duplicated into three channels as RGB inputs to MobileNetV2. The pretrained MobileNetV2 feature extractor is finetuned on three emotions: Negative, Neutral, and Positive. Experimental results on the EEG Brainwave Emotions dataset show well-converged learning and good generalization, with an overall accuracy of 84.78%, and per-class accuracies of 96.5% (Neutral), 90.8% (Negative), and 66.9% (Positive). The proposed method effectively addresses the limited feature separability in raw EEG vectors by leveraging time–frequency representations and deep CNN feature extraction. Our results verify that the EEG-to-scalogram conversion, in conjunction with LDTL, is a promising, efficient solution for real-time generalized emotion recognition and affective computing.