【摘要】 在竹材缺陷识别的研究中，竹片形状、缺陷部位颜色深浅及裂纹大小差异都是制约模型识别准确率的关键。针对上述问题，提出一种适用于中小数据集的多尺度卷积神经网络融合Transformer的竹材缺陷识别方法，以更好地提高竹材缺陷识别的准确率。该方法在卷积神经网络的主干上进行改进，从获取不同尺度语义信息的角度出发，首先利用卷积神经网络在不同尺度的特征图上捕捉图像局部语义信息，然后将不同尺度的语义特征映射为特征符号，同时引入Sinkhorn分词器对不同阶段的卷积神经网络特征符号化以减少特征冗余，再通过Transformer对特征符号之间的关系进行建模以学习图像全局语义信息。试验结果表明，与VGG16、ResNet50、DenseNet121、ViT这4种深度学习模型相比，基于多尺度卷积神经网络融合Transformer的方法能够更高效地提高竹材缺陷识别模型的性能，在竹材缺陷图像数据集上的平均识别准确率达到了99.13%。该方法识别速度更快、精度更高，且具有良好的鲁棒性，为竹材缺陷的实时自动识别提供了新思路，同时也验证了所提出方法的有效性。更多还原

【Abstract】 In the research of bamboo defect recognition, the different shapes of bamboo slices, the different color shades of defect areas, and the size of cracks are key issues that influence the accuracy of model recognition. A multi-scale convolutional neural network fused with Transformer architecture was proposed for bamboo defect recognition, particularly suitable for small and medium-sized datasets, aiming at enhancing the accuracy of defect detection. This study improved the backbone of convolutional neural networks, by focusing on capturing semantic information at various scales. Firstly, convolutional neural networks were used to capture local semantic information of images on feature maps of different scales. Then, semantic features of different scales were mapped into feature symbols. At the same time, a Sinkhorn tokenizer was introduced to symbolize the features of convolutional neural networks at different stages to reduce feature redundancy. Finally, Transformer was used to model the relationship between feature symbols to learn global semantic information of images. A series of experiments were conducted to evaluate the performance of the proposed method. This study utilized a dataset comprising 6 360 images of bamboo defects categorized into four types. The dataset was split at a ratio of 8∶2 for training and validation. During the training phase, an end-to-end training strategy was employed, optimizing model hyperparameters by minimizing the loss function. In the inference phase, the model was evaluated using the set of validation, and performance differences were compared between the proposed method and deep learning models including VGG16, ResNet50, DenseNet121 and ViT. The experimental results showed that the method based on multi-scale convolutional neural network fusion Transformer can more efficiently improve the performance of the bamboo defect recognition model compared with the four deep learning models, i.e., VGG16, ResNet50, DenseNet121 and ViT. The average recognition accuracy on the bamboo defect image dataset in this study reached 99.13%. Further results of comparative experiments revealed that the proposed model achieved an improvement in accuracy ranging from 0.16% to 7.46% compared to VGG16, ResNet50, DenseNet121 and ViT models. Additionally, it demonstrated advantages in metrics such as parameters, FLOPs and inference time. This method had faster recognition speed, higher accuracy, and good robustness, which provides a new idea for real-time automatic identification of bamboo defects, also verifies the effectiveness of the proposed method.更多还原