结合注意力与跨层双线性网络的细粒度图像分类

Fine⁃grained Image Classification Combining Attention and Cross Layer Bilinear Networks

针对细粒度图像分类任务中,对物体进行精准定位并提取更具表达力特征的难题,提出结合注意力机制与跨层特征融合的分类算法:一方面在网络中融入的混合注意力机制能够让系统忽略图像中的背景干扰信息,而将更多的关注重点放在有效信息上,以此来提高模型的判别性区域定位能力;另一方面通过跨层交互能够使得融合后得到的特征向量中包含了更多的局部特征信息,丰富了细粒度特征的学习.实验结果表明,该网络模型在公开数据集CUB200-2011上性能较好,准确率达到了88.2%,相比现有主流方法有明显提升,证明了本模型在细粒度图像分类任务上的有效性和优越性.

A classification algorithm combining attention mechanism and cross-layer feature fusion is proposed to address the challenge of pinpointing objects and extracting more expressive features in fine-grained image classification tasks.On the one hand,the hybrid attention mechanism incorporated in the network allows the system to ignore the background interference infor⁃mation in the image and focus more on the valid information to improve the discriminative region localization ability of the mod⁃el;on the other hand,the cross-layer interaction enables the fused feature vector to contain more local feature information and enrich the learning of fine-grained features.The experimental results show that the network model achieves good performance on the public dataset CUB200-2011 with an accuracy rate of 88.2%,which is a significant improvement compared with the exist⁃ing mainstream methods,and proves the effectiveness and superiority of the model for fine-grained image classification tasks.