Classification of fine-grained skin lesions based on ES-ResNet50V2

To address the current challenges of heavy workload and low efficiency in skin lesion diagnosis, as well as the uneven distribution of sample numbers in skin lesion classification images, this paper proposes the ES-ResNet50V2 algorithm. Compared with the traditional ResNet50V2 algorithm, it removes the last fully connected layer, and adds SE and ECA modules after the feature extraction phase to enhance the model's perception of key features. The SE module recalibrates channel weights through global average pooling and two fully connected layers, while the ECA module uses one-dimensional convolution to efficiently achieve channel attention. In addition, a weighted sampler is introduced to solve the problem of uneven sample distribution. The accuracy of the ES-ResNet50V2 model proposed in this paper is about 9.45% higher than that of the ResNet50V classification. It is very important to popularize the automatic skin lesion recognition tool.