| 摘要: |
| 为了解决林木业生产中存在的大批量密集堆叠木材识别和统计难题，本文提出了Wood R-DCN（Wood Region-based Deformable Convolution Network）模型。首先，构建了一个名为Wood-226的密集型堆叠的真实原木数据集，并引入了USM数据预处理技术以提高数据质量。其次，利用可变形卷积技术构建resnet的可变形残差块（DC Bottleneck），并根据提取到的特征直接产生目标检测候选框。最后，为了验证Wood R-DCN模型的有效性，在Wood-226数据集上进行了测试，对于小批量堆叠木材，模型的召回率达到了99.3%；对于大批量密集堆叠木材，召回率也能达到96%以上。在所有实验中，模型的准确率均保持在99%以上。此外，为了测试检测算法的泛化性，也在公开的钢材数据集上也进行了实验，发现召回率和准确率也能达到99%。 |
| 关键词: 木材检测；可变形卷积；Mask R-CNN；残差网络 密集堆叠检测 |
| DOI： |
| 投稿时间：2024-07-23修订日期：2024-11-10 |
| 基金项目:科技部科技创新2030-“脑科学与类脑研究”重大项目（2021ZD0201904）广西科技重大专项（桂科AA22068057） |
|
| Wood R-DCN: A Dense Stacked Wood Detection Model* |
|
QIN Xiao1, QIAN Quanmei1, LI Xiaosen2, LU Hongfei1, WANG Wenji1, PENG Lei1
|
| (1.Guangxi Key Laboratory of Human-Computer Interaction and Intelligent Decision Making,Nanning Normal University,Nanning;2.School of Artiffcial Intelligence,Guangxi Minzu University,Nanning) |
| Abstract: |
| In order to solve the identification and statistical difficulties of large-volume densely stacked logs that exist in the production of forest and timber industry, this paper proposes the Wood R-DCN (Wood Region-based Deformable Convolution Network) model. First, a densely stacked real log dataset named Wood-226 is constructed, and USM data preprocessing technique is introduced to improve the data quality. Second, a deformable residual block (DC Bottleneck) of resnet is constructed using deformable convolution technique and target detection candidate frames are directly generated based on the extracted features. Finally, in order to validate the effectiveness of the Wood R-DCN model, it was tested on the Wood-226 dataset, and the model's recall reached 99.3% for small quantities of stacked wood, and over 96% for large quantities of densely stacked wood. In all experiments, the accuracy of the model remained above 99%. In addition, in order to test the generalization of the detection algorithm, experiments were also conducted on the publicly available steel dataset as well, and it was found that the recall and accuracy could also reach 99%. |
| Key words: wood detection, deformable convolution, Mask R-CNN, residual network, dense stacked detection |
