基于B-CNN算法的汽车数据集细粒度图像分类分析

Fine-Grained Image Classification Analysis of Automotive Dataset Based on B-CNN Algorithm

为了提高计算机视觉领域细粒度分类方法性能,采用双线性卷积神经网络(B-CNN)对其进行建模研究。以StanfordCars汽车数据集为研究对象,开展细粒度图像分类分析。对B-CNN进行阐述。应用了组归一化、dropout技巧,调整可训练参数,并在汽车数据集上进行试验。比较了经典卷积神经网络和B-CNN。在网络训练中,合适的学习率有助于提升训练速度。在损失函数变化幅度变小到一定范围时,将学习率变小,可跳出局部最优解,寻找全局最优解,避免陷入死循环。结果显示,与单路神经网络相比,B-CNN在输入图像大小为(224,224)时,在汽车数据集中的准确率提升了16%。B-CNN适用于一些细粒度图像分类任务,能提升分类准确率,具有很好的实际应用效果。

To improve the performance of fine-grained classification methods in the field of computer vision,a bilinear convolutional neural network(B-CNN)is used to model the study.The StanfordCars automotive dataset is used as the study object to carry out fine-grained image classification analysis.The B-CNN is elaborated.Group normalization and dropout techniques are applied,trainable parameters are adjusted,and experiments are conducted on the car dataset.Classical convolutional neural networks and B-CNN are compared.In the network training,a suitable learning rate helps to increase the training speed,and can reduce the learning rate when the magnitude of the loss function variation becomes small to a certain range can jump out of the local optimal solution to find the global optimal solution and avoid getting into a dead loop.The results show that compared with the single-way neural network,the accuracy of B-CNN in automotive dataset improves by 16%when the input image size is(224,224).B-CNN is suitable for some fine-grained image classification tasks,which can improve the classification accuracy and has good practical application.