基于跨层精简双线性网络的细粒度鸟类识别

细微的类间差异和显著的类内变化使得细粒度图像分类极具挑战性。为了对鸟类图像进行细粒度识别,提出一种基于跨层精简双线性池化的深度卷积神经网络模型。首先,根据Tensor Sketch算法计算出多组来自不同卷积层的精简双线性特征向量;其次,将归一化后的特征向量级联送至softmax分类器;最后,引入成对混淆对交叉熵损失函数进行正则化以优化网络。提出的模型无需额外的部件标注,可进行端到端的训练。结果表明,在公开的CUB-200—2011鸟类数据集上,该模型取得了较好的性能,识别正确率为86. 6%,较BCNN提高2. 5%。与多个先进细粒度分类算法的对比,验证了提出模型的有效性和优越性。

Fine-grained Bird Recognition Based on Cross-layer Compact Bilinear Network

Fine-grained image classification are quite challenging due to subtle inter-class differences and large intra-class variations. In order to fine-grain the bird image, a deep convolutional neural network model based on cross-layer compact bilinear pooling was proposed. Firstly, multiple sets of compact bilinear feature vector from different convolutional layers were calculated according to the Tensor Sketch algorithm. And then, the normalized feature vectors were concatenated and sent to the softmax classifier. Finally, the pairwise confusion was used to regularize the cross entropy loss function in order to optimize network. The model can be end-to-end trained without additional part annotation. The result shows that it achieves good performance on the public CUB-200-2011 bird dataset and gets 86.6% recognition accuracy, which surpasses BCNN by 2.5%. The comparison with several advanced algorithms proves the validity and superiority of the proposed model.