As one of the most prevalent cancers in the world, skin cancer requires early and precise detection to improve clinical outcomes. The expertise of dermatologists is the foundation of traditional diagnostic techniques, which can be laborious and subjective. With Convolutional Neural Networks (CNNs) being particularly good at extracting local features and Vision Transformers (ViTs) capturing global dependencies, recent developments in deep learning have shown great promise in automated skin cancer classification. Nevertheless, the absence of integrated local-global feature learning limits the lesion classification capabilities of current standalone architectures. This paper suggests a CNN-Transformer Hybrid Model, which combines Vision Transformers for self-attention-based global feature learning with ResNet-50/EfficientNet for CNN-based feature extraction. The Synthetic Minority Over-Sampling Technique (SMOTE), data augmentation, and image normalization are all part of a comprehensive preprocessing pipeline that is used to improve model generalization. The Transformer module applies multi-head self-attention to refine the features derived from the CNN, followed by a fusion strategy that enhances lesion classification performance. The model demonstrates superior performance over existing state-of-the-art methods in both precision and recall, attaining an accuracy of 98.7% on the ISIC 2019 dataset. The findings show that a clinically feasible method for automated and effective skin cancer detection is offered by the suggested framework.