This study proposes the detection of skin cancer through dermoscopic images, using an ensemble of deep neural network classifiers. The proposed method does not require image pre-processing techniques. Public datasets such as HAM10000 and ISIC (2016 and 2017) are used. Due to the unbalanced nature of the classes, data augmentation, sample balancing and weight balancing strategies are adopted. Pre-trained neural network architectures such as DenseNet201, ResNet101 and MobileNetV2 are evaluated individually and as an ensemble, improving robustness and diagnostic accuracy. The accuracy of the classification of lesions into three classes (melanoma, non-melanoma cancer and benign lesions) was 92.33% and the binary classification (benign and malignant lesions) was 89.50%. The results of the binary classification outperformed the state of the art. This work demonstrates the relevance of deep learning techniques in improving diagnostic accuracy for skin cancer, with a positive impact on early detection and treatment.