基于部件关注DenseNet的细粒度车型识别

针对细粒度车型识别率低,车型区别主要集中在鉴别性部件上以及深度学习不能有效对部件进行关注的问题,提出一种基于部件关注DenseNet(part-focused DenseNet, PF-DenseNet)的细粒度车型识别模型。该模型可以基于细粒度车型的车灯、车标等区分性部件进行有效分类,通过处理层(process layer)对车型部件信息反复加强提取并进行最大池化下采样,获取更多的车型部件信息,然后通过密集卷积对特征通道进一步复用提取,密集卷积前嵌入独立组件(independent component, IC)层,获得相对独立的神经元,增强网络独立性,提高模型的收敛极限。实验结果表明,该模型在Stanford cars-196数据集上的识别准确率、查全率和F₁分别达到95.0%、94.9%和94.8%,高于经典卷积神经网络,并具有较小的参数量,与其他方法相比实现了最高准确率,验证了该车型识别模型的有效性。

Fine-grained vehicle-type identification based on partially-focused DenseNet

Given that fine-grained model recognition rates are low and are mainly concentrated in the diagnostic model difference between parts and that deep learning cannot effectively concern parts, we put forward a fine-grained recognition model—the partially-focused DenseNet. The model can be classified effectively based on its discriminative parts, such as lights and marks of fine-grained vehicle models. First, through the Process Layer, the vehicle part information is repeatedly extracted, and the maximum pool sampling is carried out to obtain more vehicle part information. Then, feature channels are further extracted by multiplexing through dense convolution, and the independent component layer is embedded before dense convolution to obtain relatively independent neurons. This enhances network independence and improves the convergence limit of the model. Experiments show that the model’s recognition accuracy, recall rate, and F₁ on the Stanford cars-196 data set reach 95.0%, 94.9%, and 94.8%, respectively, which are higher than the classic convolutional neural network and have a smaller number of parameters. Compared with other methods, the highest accuracy rate is achieved, verifying the effectiveness of the vehicle recognition model.