基于改进注意力机制和多语义特征增强的自然环境下枣品种识别方法

Jujube Variety Recognition Based on Improved Attention Mechanism and Multi-semantic Feature Enhancement

摘要:

针对目前自然环境下枣品种识别准确率较低的问题，提出了一种基于注意力机制和多语义特征增强的枣品种识别模型（ICBAM_MSFE_Res50）。该模型在ResNet-50基础上，引入改进注意力机制（Improved convolutional block attention module， ICBAM），ICBAM采用一维卷积和多尺度空洞卷积对卷积块注意力模块（CBAM）进行改进，消除了特征图降维时的信息损失，降低了模型计算量和参数量，提高了模型对枣果区域细粒度特征的提取能力。同时，提出了多语义特征增强（Multi semantic feature enhancement，MSFE）模块，该模块通过枣果区域定位算法提取更多枣果局部显著特征，并采用显著性特征抑制算法迫使模型学习枣果次要特征，从而达到枣果多种语义特征学习。实验结果表明，在20类枣品种数据集上，本文模型准确率为92.20%，与ResNet-50相比，提高4.26个百分点。对比AlexNet、VGG-16、ResNet-18、InceptionV3模型，准确率分别提高15.84、9.22、6.86、3.55个百分点。对比其他枣品种识别方法，本文方法在20种枣品种识别中表现最优，可为自然环境下枣品种识别研究提供参考。

Abstract:

In response to the low accuracy of jujube variety recognition in current natural scenarios, a jujube variety recognition model was proposed based on attention mechanism and multi-semantic feature enhancement (ICBAM_MSFE_Res50). On the basis of ResNet-50, the attention mechanism ICBAM (improved convolutional block attention module) was introduced. ICBAM improved the convolutional block attention module (CBAM) by using one-dimensional convolution and multi-scale hole convolution, eliminating information loss during feature map dimensionality reduction, reducing the computational and parameter complexity of the model, and improving the model’s ability to extract fine-grained features in jujube fruit regions. At the same time, a multi-semantic feature enhancement (MSFE) module was proposed, which extracted more local salient features of jujube fruit through jujube fruit region localization algorithm, and used saliency feature suppression algorithm to force the model to learn secondary features of jujube fruit, thereby achieving the learning of multiple semantic features of jujube fruit. The experimental results showed that the accuracy of the model on the dataset of 20 types of jujube varieties was 92.20%, which was 4.26 percentage points higher than that of ResNet-50. Compared with the AlexNet, VGG-16, ResNet-18, and InceptionV3 models, the accuracy was improved by 15.84, 9.22, 6.86, and 3.55 percentage points, respectively. Compared with other jujube variety recognition methods, this method still performed the best in the recognition of 20 types of jujube, which can provide reference for research on jujube variety recognition in natural scenarios.