Abstract: Existing deep learning methods still face the following challenges in the task of high-precision imperfect grain recognition: the key discriminative features of imperfect kernels are often distributed in image regions with varying scales and random positions, making it difficult to perceive these regions stably and comprehensively; meanwhile, the fine-grained features of multiple imperfect grain categories exhibit diverse expression patterns, which a unified modeling path struggles to optimize simultaneously. To address these issues, a globally guided two-stage local feature learning framework was proposed based on the TransNeXt architecture. Deep representations of key discriminative regions were extracted under holistic perception and further refined through fine-grained modeling. Additionally, independently optimized network branches were designed for different categories, with all branches sharing the backbone to enable efficient adaptation and lightweight scalability. A dataset of imperfect grain kernels was also constructed, covering multiple grain varieties and annotated with standardized categories and region-level labels. Experimental results show that an accuracy of 99.62% was achieved on the test set, representing a 0.14 percentage point improvement over TransNeXt-Micro, which verifies the effectiveness and scalability of the proposed framework in complex fine-grained image recognition tasks.

Key words: imperfect grain recognition, TransNeXt, fine-grained image recognition, local feature learning, multi-branch network

摘要： 现有深度学习方法在高精度粮食不完善粒识别任务中仍面临以下问题：不完善粒的关键判别特征局部分布于尺度不一、位置随机的图像区域，模型难以稳定、全面地感知这些区域；多类不完善粒的细粒度判别特征表达形式多样，统一建模路径难以同时优化所有类别的识别性能。为此，基于TransNeXt架构提出了一种全局引导的两阶段局部特征学习框架，在整图感知的基础上提取关键判别区域的深层表征，并在该区域上完成细粒度特征的建模。进一步针对不同类别构建独立优化的网络分支，多个分支共享主干网络，实现对多类目标的高效适配与轻量扩展。为支撑上述方法，构建了一个覆盖多粮食品种、具备标准化类别与判别区域位置标注的不完善粒数据集。实验结果表明，所提方法在测试集上达到了99.62%的准确率，较TransNeXt-Micro提升0.14个百分点，验证了其在复杂细粒度图像识别任务中的有效性与可扩展性。

关键词: 不完善粒识别, TransNeXt, 细粒度图像识别, 局部特征学习, 多分支网络