Early Melanoma detection is a challenging task due to an imbalanced dataset and less number of labeled images. There are many methods based on different Deep CNN architectures however as the interclass similarity is very low, more focus on the target area is highly required. In this paper, a dual Attention module is proposed at the skip connection part of ResNet101 after the convolution block. The proposed model used a channel attention module to increase the weight of the channel that contains more relevant information. The model also used a spatial attention module to highlight local features in the spatial domain. Proposed model score classification accuracy 95.1%, sensitivity 87.6%, and specificity 95.8%. The proposed model achieves better classification accuracy, sensitivity, and specificity than other state-of-the-art methods.