Improved AlexNet Based Recognition Algorithm for Deformable Convolution Skin Lesions

In order to solve the problem of poor adaptability of standard convolution to unknown changes and high similarity between classes of skin lesions, large differences within a class, and serious class imbalance in data, a skin lesion recognition algorithm of deformable convolutional network was proposed based on improved AlexNet. First, an improved AlexNet deformable convolutional network model was established, adding sampling offset, making the position of the convolution kernel sampling point change adaptively according to the image content. And it was arranged to automatically adjust different scales or acceptable information field, and to extract finer features than standard convolution. And then, a weighted loss function of the cross-entropy loss function and the focus loss function was used to reduce the weight of samples classified easily in training, making the model focus on difficult and error-prone samples to solve the problem of sample proportion imbalance and to optimize the recognition rate of the model. Finally, a simulation experiment was carried out with the HAM10000 data set. The subjective and objective experimental results show that the proposed method is superior to the existing methods in identifying seven skin lesions, and has higher accuracy, specificity and robustness.