Skin cancer is a main public health issue, as is the most prevalent form of cancer, shows more than half of all cancers diagnosed worldwide. In this study, a novel deep learning-based framework is developed for the classification of skin lesions. The primary objectives of this research are to reach a high classification rate on a public skin cancer image dataset and use the effectiveness of the different deep feature generators together. The presented model uses five pre-trained deep learning models as feature generators. By generating 1000 features from each model, 5000 features are extracted totally. In the features selection phase, an iterative and improved version of the minimum redundancy maximum relevance (mRMR) feature selection technique, called ImRMR, is used. The proposed hybrid deep features extraction and ImRMR based feature selection model reached 96.58% accuracy on the used dataset. The calculated results were also compared to other state-of-the-art skin cancer detection models, and the proposed model achieved higher results than the previous works.