自适应随机森林迁移学习的路面状态检测

Road condition detection based on adaptive random forest migration learning

为提升少量数据集条件下路面状态检测的准确度和效率,提出基于改进卷积神经网络模型VGG16和自适应随机森林迁移学习的路面图像分类方法 .对改进VGG16模型进行迁移学习,将大数据集训练得到的VGG16网络卷积层、池化层、全连接层进行迁移,采用随机森林分类算法代替VGG16网络的softmax层进行重新学习训练,解决softmax强调特征之间独立性的缺点.此外,改进量子狼群算法的量子旋转门更新策略,将其用于随机森林超参数优化,保证随机森林以最佳的参数进行迁移学习训练,进一步提升模型泛化能力.实验结果表明,在自建以及Kaggle网站提供的图像分类实验中,图像识别精度为98.08%,分类速度也得到显著提升.

In order to improve the accuracy and efficiency under small datasets,a road image classification method based on improved convolutional neural network model VGG16 and adaptive random forest transfer learning is proposed.Transfer learning was carried out on the improved VGG16 model,and the convolutional layer,pooling layer,and fully connected layer of the VGG16 network trained on a large dataset were trans-ferred.The random forest classification algorithm was used to replace the softmax layer of the VGG16 network for re-learning and training,overcoming the disadvantage of softmax’s emphasis on the independence between features.In addition,the quantum revolving door update strategy of the quantum wolf pack algorithm is im-proved and has been used for the random forest hyperparameter optimization to ensure that the random forest is transferred with the best parameters to train and further improve the generalization ability of the model.The ex-perimental results show that in the image classification experiments which is self-built and provided by Kaggle website,the image recognition accuracy is 98.08%,and the classification speed has also been significantly im-proved.