基于深度特征局部重采样融合的多种类水稻种子识别

Multi-type Rice Seed Recognition Based on Local Resampling Fusion of Deep Features

摘要:

针对多种类水稻种子识别过程中，形态特征较多、分类难度较大的问题，本文提出了一种基于深度特征局部重采样融合（Depth feature local resampling fusion，DFLRF）的分类网络，对36种水稻种子进行分类识别。首先，该方法使用ConvNeXt作为骨干网络提取水稻种子特征；其次，采用特征强化注意力模块（Feature intensification attention module，FIAM）构造全局特征采集分支，使用多通道卷积局部重采样模块（Multi-channel convolutional local resampling module，MCLRM）和FIAM构建局部特征采集分支；最后，将输出的全局特征和局部特征进行融合，在CosFace损失约束下准确识别出具有近似特征的不同种类水稻种子。本研究使用自采数据集，实验得出，新模型ConvNeXt-DFLRF总体准确率达到86.90%，较基础模型提高5.88个百分点，与InceptionResNetV2和EfficientNetV2等主流模型相比，总体识别准确率提升2.92~8.80个百分点，整体识别效果最优。本文所提出模型能够有效地对36种水稻种子进行分类，为多种类水稻种子分类识别的研究提供了一种新颖且有效的方法。

Abstract:

In the process of using rice seeds, it was essential to classify them, but this problem was highly challenging because rice seeds have similar visual appearance. To address the challenges associated with recognizing multiple rice seed varieties, specifically the abundance of morphological features and the high classification complexity, a classification network based on depth feature local resampling fusion (DFLRF) was proposed to classify and identify 36 types of rice seeds. ConvNeXt was adopted as the backbone network to extract feature representations of the seeds. Subsequently, a global feature extraction branch was constructed by using the feature intensification attention module (FIAM), while the local feature extraction branch was built by using both the multi-channel convolutional local resampling module (MCLRM) and FIAM. The global and local features were then fused, and under the supervision of the CosFace loss, rice seed varieties with highly similar visual characteristics were accurately identified. A self-collected dataset was utilized. Experimental results demonstrated that the proposed ConvNeXt-DFLRF model achieved an overall accuracy of 86.90%, marking a 5.88 percentage points improvement over the baseline model. When compared with mainstream architectures such as InceptionResNetV2 and EfficientNetV2, the proposed model exhibited improvements ranging from 2.92 to 8.8 percentage points, achieving the highest recognition performance among the tested methods. These findings confirmed that the proposed model was effective in classifying 36 rice seed varieties and provided a novel and efficient solution for multi-type rice seed classification and recognition tasks.