Low quality dermal image classification using transfer learning

摘要

In this study, we investigate three class skin lesion classification problem of a low quality and small size dataset using transfer learning using AlexNet deep Convolutional Neural Network (CNN). Our approach involves modifying the pre-trained AlexNet model; through replacing the decision layer to be compatible with our three class problem. In addition, we propose adding two dropout layers to overcome the over fitting problem. The fine tuning process of the complete network, based on stochastic gradient descent, is performed using skin lesion dataset. Furthermore, we investigated augmenting the original dataset through three flipping directions and sixteen rotation angles processes using a new methodology. The proposed algorithm has been compared with a hand crafted features, based on Local Binary Pattern (LBP) representation followed by Support Vector Machine (SVM) classifier. Increasing the dataset size has dramatically boosted the performance of classifiers achieving accuracy of 98.67% for the modified AlexNet compared to 96.8% using the LBP based system.