Jiang Yuying, Chen Xinyu, Ge Hongyi, Wen Xixi, Jiang Mengdie, Zhang Yuan
Key Laboratory of Grain Information Processing & Control, Ministry of Education, Henan University of Technology, Zhengzhou, China.
Henan Provincial Key Laboratory of Grain Photoelectric Detection and Control, Zhengzhou, China.
Front Plant Sci. 2025 Jun 4;16:1490384. doi: 10.3389/fpls.2025.1490384. eCollection 2025.
Wheat is susceptible to mold growth due to storage conditions, which subsequently affects its quality; therefore, timely and rapid identification of moldy wheat is critically important. In order to achieve high-precision recognition and class classification of wheat with different degrees of mold, a multi-scale context and feature pyramid based moldy wheat recognition network (MSCFP-Net) is proposed. Firstly, the network uses the residual network ResNeXt as the baseline network, and incorporates a multi-scale contextual feature extraction module, which is more helpful to determine the important discriminative regions in the whole image to extract more image detail features. In addition, a coordinated attention mechanism module is introduced to perform global average pooling from both directions to learn the importance of different regions in the input features in a dynamically weighted manner. Moreover, a bidirectional feature pyramid network is embedded into the baseline model, so that certain coarse-grained features and fine-grained features are retained in the processed output features at the same time to improve the network recognition accuracy. Compared with the baseline network, the four evaluation indexes of Accuracy, Precision, Recall and F1-Score of MSCFP-Net are improved by 1.08%, 1.25%, 0.53% and 0.91%, respectively. In addition, a series of comparison experiments and ablation experiments show that the classification network constructed in this paper has the best fine-grained classification performance for moldy wheat THz images.
由于储存条件，小麦容易发生霉菌生长，这随后会影响其品质；因此，及时快速地识别发霉小麦至关重要。为了实现对不同霉变程度小麦的高精度识别和类别分类，提出了一种基于多尺度上下文和特征金字塔的发霉小麦识别网络（MSCFP-Net）。首先，该网络使用残差网络ResNeXt作为基础网络，并融入多尺度上下文特征提取模块，这更有助于确定整个图像中的重要判别区域，以提取更多图像细节特征。此外，引入了一种协同注意力机制模块，从两个方向进行全局平均池化，以动态加权的方式学习输入特征中不同区域的重要性。而且，将双向特征金字塔网络嵌入到基础模型中，使得在处理后的输出特征中同时保留一定的粗粒度特征和细粒度特征，以提高网络识别准确率。与基础网络相比，MSCFP-Net的准确率、精确率、召回率和F1分数这四个评估指标分别提高了1.08%、1.25%、0.53%和0.91%。此外，一系列对比实验和消融实验表明，本文构建的分类网络对发霉小麦太赫兹图像具有最佳的细粒度分类性能。
