【摘要】 为精准识别并高效分割粘连烟丝，提出一种基于改进掩码区域卷积神经网络模型(Mask R-CNN)的粘连烟丝识别方法。首先，采集粘连烟丝图像，通过图像增强操作将数据集扩充到训练模型所需的样本数量；其次，在Mask R-CNN模型的基础上对训练样本中的粘连烟丝图像进行边缘特征提取、分形特征提取，获得更清晰、连续的图像边缘特征信息和纹理特征信息；再次，将原始特征、边缘特征、分形特征进行融合以充分利用不同层次的特征信息，丰富底层特征；最后，通过引入混合注意力机制关注特征图的通道和空间维度，从而提高粘连烟丝识别的效率和准确性。模型性能对比结果表明：基于改进Mask R-CNN模型的识别方法的平均交并比(Avg.MIoU)为85.29%,类别平均像素准确率(Avg.MPA)为84.33%,其能够快速、准确地识别并分割出单根烟丝，识别效果优于Mask R-CNN和DeepLabV3+模型识别方法，可为后续烟丝宽度检测提供技术支持。更多还原

【Abstract】 To achieve accurate identification and efficient segmentation of adhesive tobacco shreds, a method for adhesive tobacco shreds recognition based on an improved Mask R-CNN(Mask Region-based Convolutional Neural Network) model was proposed. Firstly, adhesive tobacco shreds images were collected, and the dataset was augmented through image enhancement operations to expand it to the required sample size for training the model. Secondly, edge feature extraction and fractal feature extraction were performed on the adhesive tobacco shreds images in the training set based on Mask R-CNN, resulting in clearer and more continuous image edge features and texture feature information. Subsequently, the original features, edge features, and fractal features were fused to fully utilize features at different levels and enrich low-level features. Finally, by introducing a hybrid attention mechanism that focused on both channel and spatial dimensions of feature maps, the efficiency and accuracy of tobacco shred recognition were improved. Experimental results showed that the mean intersectionover union(Avg.MIoU) of the recognition method based on the improved Mask R-CNN model was 85.29%, and the mean class pixel accuracy(Avg.MPA) was 84.33%, under different adhesion conditions enabling precise identification of tobacco shreds and outperforming the original Mask R-CNN and DeepLabV3+ models. This method could rapidly and accurately identify and segment adhesive tobacco shreds, providing technical support for subsequent tobacco shred width detection.更多还原