基于跨粒度特征渐进融合的细粒度锁芯孔识别

Fine-Grained Lock Cylinder Hole Recognition Based on the Progressive Fusion of Cross-Granularity Features

针对细粒度图像类间差异小、鉴别性特征难以捕捉、识别精度低等问题,提出一种基于跨粒度特征渐进融合的细粒度图像识别方法。首先,使用随机区域混淆模块(RRCM)生成不同粒度级别的图像,用于训练骨干网络ConvNeXt的不同阶段;其次,使用随机样本交换模块(RRSM)增强不同粒度图像在模型中层的表征;然后,使用渐进式多粒度训练策略、互信道损失函数进行模型训练,协同融合跨粒度信息;最后,拼接融合多粒度特征并组合分类器,获得最终识别结果。实验结果表明,所提方法在CUB-200-2011、Stanford Cars和FGVC-Aircraft等3个公开数据集上的识别精度分别为92.8%、95.5%和94.0%,优于当前主流的细粒度图像识别方法。在自行构建的Lock-Hole锁芯孔数据集上的识别精度达到97.3%,单张图像平均识别时间为0.016 s,能够实现锁芯孔图像的精准识别,满足应急开锁场景下的快速锁芯孔识别要求。

A fine-grained image recognition method based on the progressive fusion of cross-grained features is proposed to address the problems of small differences between classes of fine-grained images,difficulty in capturing discriminative features and low recognition accuracy.First,the random region confusion module is used to generate images with different granularity levels for training various stages of the backbone network ConvNeXt.Second,the image representation of differing granularity in the middle layer of the model is enhanced using the random sample-swapping module.Then,the progressive multigranularity training strategy and mutual belief channel loss function are used for model training to fuse cross-granular information collaboratively.Finally,to obtain the final recognition results,the multigranularity features are integrated and fused to combine the classifiers.The experimental results demonstrate that the recognition accuracies of this method on three public datasets are 92.8%(CUB-200-2011),95.5%(Stanford Cars),and 94.0%(FGVC-Aircraft),which are better than the current mainstream fine-grained image recognition methods.The recognition accuracy on the selfconstructed Lock-Hole dataset reaches 97.3%,and the average recognition time of a single image is 0.016 s,which can realize the accurate recognition of the lock-hole image and satisfy the requirement of fast lock-hole recognition in emergency unlocking scenarios.