融合通道与位置信息的ResNet细粒度图像识别

ResNet fine-grained image identification with fused channel and location information

在细粒度视觉识别(FGVR)领域,由于高度近似的类别之间差异细微,因此图像细微特征的精确提取对识别的准确率有着至关重要的影响。针对该问题,提出了融合通道与位置信息的残差网络(ResNet)细粒度图像识别算法。首先,通过引入超轻量化空间与位置感知注意力模块(ULSLAM)改进的ResNet细粒度图像特征提取网络,提高了细粒度图像特征尺度丰富性与多样性且有效增强了上下文特征非线性依关系;其次,使用融合通道与位置信息特征学习网络,利用权重方差度量获得特征提取网络显著特征以馈送到识别器进行最终有效识别,而后通过抑制因子抑制显著特征用于下阶段特征提取网络对细微特征进行提取。实验结果表明,该算法在数据集CUB-200-211上达到89.60%的top1准确率、98.65%的top5准确率;在数据集Stanford Cars上达到94.93%的top1准确率、98.93%的top5准确率;在FGVC-Aircraft数据集上达到93.80%的top1准确率、98.20%的top5准确率。

In the field of fine-grained visual recognition(FGVR),due to the subtle differences between highly similar categories,the accurate extraction of image fine features has a crucial impact on the accuracy of recognition.To solve this problem,a ResNet fine-grained image recognition algorithm based on fusion of channel and location information is proposed.First,by introducing the ResNet fine-grained image feature extraction network improved by Attention Component(ULSLAM),the feature scale richness and diversity of fine-grained images are improved and the nonlinear dependency of context features is effectively enhanced.Secondly,the fusion channel and the location information feature learning network are used to obtain the salient features of the feature extraction network using the weight variance measurement to feed them to the recognizer for final effective recognition,and then the salient features are suppressed by the suppression factor for the next stage feature extraction network to extract the subtle features.The experimental results show that the algorithm achieves 89.60%top1 accuracy and 98.65%top5 accuracy on CUB-200-211 dataset;The top 1 accuracy rate of 94.93%and the top 5 accuracy rate of 98.93%are achieved on Stanford Cards;On the FGVC Aircraft dataset,the top 1 accuracy is 93.80%,and the top 5 accuracy is 98.20%.