基于ResNet50改进模型的图像分类研究

Research on image classification based on ResNet50 improved model

针对深度学习中残差网络ResNet50存在的信息丢失、特征提取不充分、网络过拟合和训练困难等问题,文中提出一种基于改进ResNet50的图像分类算法。针对残差网络ResNet50在提取特征时存在丢失输入特征映射情况,造成信息丢失的问题,对主干网络中Stage4的下采样块添加平均池化层,进一步提高网络特征提取能力;针对ResNet50训练过程中存在网络过拟合以及泛化能力差的问题,使用标签平滑方法对交叉熵损失函数进行修改,有效缓解网络损失值震荡幅度;针对ResNet50计算量大、训练困难的问题,使用混合精度和余弦退火衰减方法对模型进行训练,在加快网络收敛速度的同时提高模型的分类精度。实验结果表明,与原ResNet50网络相比,文中算法在ImageNet-1k数据集上Top1和Top5的精度分别提升3.2%和1.6%,能够更好地应用于图像分类任务。

In allusion to the problems of information loss,insufficient feature extraction,over network fitting and training difficulties of the residual network ResNet50 in depth learning,an image classification algorithm based on the improved ResNet50 is proposed. In order to solve the problem of information loss caused by the loss of input feature mapping in the feature extraction of residual network ResNet50,the average pooling layer is added to the down-sampling block of Stage4 in the backbone network to further improve the network feature extraction ability. In view of the problems of over fitting and poor generalization ability of the network in the ResNet50 training process,the label smoothing method is used to modify the cross entropy loss function,which can effectively alleviate the oscillation amplitude of the network loss value. In allusion to the large computed amount and difficult training of ResNet50,the hybrid precision and cosine annealing attenuation method are used to train the model,so as to improve the classification accuracy of the model while accelerating the convergence speed of the network. The experimental results show that in comparison with the original ResNet50 network,this algorithm can increase the accuracy of Top1 and Top5 on ImageNet-1k dataset by 3.2% and 1.6%,respectively,which can be better applied to image classification.