基于可变形卷积AlexNet与软注意力机制的皮肤病变识别算法

Skin Disease Identification Algorithm Based on Deformable Convolutional AlexNet and Soft Attention Mechanism

摘要: 为解决皮肤病识别领域中数据集类别不平衡、模型复杂度高以及准确率低的问题，提出了一种基于可变形卷积AlexNet与软注意力机制的皮肤病变识别算法。首先，提出改进的可变形卷积AlexNet网络模型，提高模型辨析力的同时，降低了模型的参数量，加快模型的训练和测试效率。然后，在改进的模型中集成了软注意力机制，使模型聚焦于皮肤病的关键特征区域，优化模型的特征提取和识别能力。最后，提出了一种联合损失函数，对焦点损失函数与交叉熵损失函数进行加权，聚焦于困难样本和易出错样本，解决因数据集类别不平衡而导致的网络朝着错误方向收敛的问题。在公开数据集进行实验，主观和客观的实验结果表明，提出算法在七种不同类别的皮肤病识别准确率高于对比算法，具有较强的鲁棒性和泛化能力。

Abstract: To address the issues of dataset class imbalance, high model complexity, and low accuracy in the field of skin disease identification, this paper proposes a skin lesion recognition algorithm based on deformable convolutional AlexNet and soft attention mechanisms. First, an improved deformable convolutional AlexNet network model is introduced, which enhances the model’s discriminative power while reducing the number of model parameters, thereby speeding up the model’s training and testing efficiency. Subsequently, a soft attention mechanism is integrated into the improved model, focusing the model on key feature areas of skin diseases to optimize its feature extraction and recognition capabilities. Finally, a joint loss function is proposed, which applies weights to the focal loss function and cross-entropy loss function, concentrating on difficult and error-prone samples to solve the issue of network convergence in the wrong direction due to dataset class imbalance. Experiments on public datasets, with both subjective and objective results, demonstrate that the proposed algorithm achieves higher accuracy in identifying seven different categories of skin diseases compared to benchmark algorithms, exhibiting strong robustness and generalizability.