基于Inception深度残差网络的皮肤黑色素癌图像分类算法

由于皮肤黑色素癌图像存在类内差异大、样本数据集小等特点, 采用深度残差网络可以有效解决训练过程中过拟合问题, 提高识别准确率. 但是深度残差网络模型的训练参数多, 时间复杂度高. 为了提高训练效率, 提高识别准确率, 首先从理论上分析了深度残差网络模型的结构, 通过修改网络结构, 利用Inception结构代替残差网络中的卷积层、池化层, 减少模型的训练参数数量, 降低时间复杂度. 在此基础上, 提出了基于Inception深度残差网络皮肤黑色素癌分类识别算法(Inception Deep Residual Network, IDRN), 用Inception结构代替残差网络中的卷积池化层, 用SeLU激活函数代替传统的ReLU函数. 之后, 在公开的黑色素癌皮肤镜图像ISIC2017数据集上进行实验验证. 理论和实验表明, 与传统的卷积神经网络ResNet50相比, 本文提出的新的分类算法降低了时间复杂度, 提高了识别准确率.

Skin Melanoma Classification Algorithm Based on Inception Deep Residual Network

Since skin melanoma images are featured by large intraclass differences and small sample datasets, the deep residual network can effectively solve the problem of over-fitting during training and improve the recognition accuracy. However, the network model has many training parameters and high time complexity. To improve the training efficiency and the recognition accuracy, we theoretically analyze its structure. By modifying the network structure, we replace the convolutional and pooling layers in the residual network with the Inception structure to lower the number of training parameters and the time complexity of the model. On this basis, we propose an Inception Deep Residual Network (IDRN) based classification and recognition algorithm for skin melanoma, where the Inception structure and the SeLU activation function respectively replace the convolutional and pooling layers and the traditional ReLU function. Subsequently, experimental validation is carried out on the published ISIC2017 dataset of dermoscopic images of melanoma. The theoretical and experimental results show that compared with the traditional convolutional neural network ResNet50, the proposed algorithm reduces time complexity and improves recognition accuracy.