基于改进可变形-端到端目标检测模型的竹片缺陷检测方法  

Defect Detection Method for Bamboo Slice Based on Improved Deformable-DETR

作　　者：马良城 徐筱茹 伍希志 MA Liangcheng;XU Xiaoru;WU Xizhi(College of Materials Science and Engineering,Central South University of Forestry and Technology,Changsha 410000,China)

机构地区：[1]中南林业科技大学材料科学与工程学院,长沙410000

出　　处：《森林工程》2025年第2期349-359,共11页Forest Engineering

基　　金：湖南省科技特派员服务乡村振兴(2023NK4285);中国博士后科学基金资助(2021M690768)。

摘　　要：目前,虽然已经有一些基于图像处理技术的竹片缺陷检测方案,但这些方案检测存在种类较少、实用性较差且难以部署在机器上等缺陷,为此,提出一种改进的竹片缺陷检测模型。该模型为改进的可变形-端到端目标检测(Deformable-DETR)模型,首先将骨干网络替换成由DCNv3卷积为核心而堆叠设计的InternImage,该网络在保留卷积神经网络(CNN)先验特性的情况下还能捕捉到长距离依赖,使提取到的特征空间语义更丰富;然后在特征提取后新增一个采样模块,该采样模块将图像特征抽象为精细的前景特征和少量粗糙的背景特征,不仅能去除冗余的背景特征信息,还能提取高语义前景信息;最后引入一种新颖的协作混合分配训练策略,该策略通过训练由一对多标签分配监督的多个并行辅助头,提高编码器在端到端检测器中的学习能力。此外,使用数据增强来扩展数据集,并使用迁移学习,以增强竹片缺陷的检测。试验结果表明,该改进方法可以提高模型的缺陷特征提取与解析的能力,并在测试数据集上取得了85.7%mAP50(全类平均精确度),单张图片推理时间为0.28 s,检测精度优于其他主流目标检测模型,为竹片缺陷检测提供新的方法。Currently,although there are some bamboo slice defect detection schemes based on image processing techniques,these schemes detect fewer types of defects,are less practical,and are difficult to deploy on machines.For this reason,an improved defect detection model for bamboo slice is proposed.Therefore,we propose an improved model for bamboo slice defect detection.The model proposed in this paper is an improved Deformable-DETR model,which firstly replaces the original backbone extraction network ResNet with InternImage,which is stacked with DCNv3 convolution as the core.This network retains the a priori properties of the traditional CNN and captures the long-range dependencies,making the extracted feature spatial semantics richer.Then,after feature extraction,a new sampling module is added,which abstracts the image feature mapping into fine a fine foreground target feature vectors and a small number of coarse background context feature vectors,which can not only remove redundant backgroud feature information but also extract high-semantic foreground.Finally,a novel collaborative hybrid allocation training scheme is introduced,which supervises the training of multiple parallel auxiliary heads through one-to-many label allocation,to easily improve the encoder′s learning capability in an end-to-end detector.In addition,data augmentation is used to extend the dataset and migration learning is used to enhance the detection of bamboo slice defects.The experimental results show that the method proposed in this paper improves the defective feature extraction and parsing ability of the model,achieves 85.7%of mAP50 on the test dataset,the inference time for a single image is 0.28 seconds,and the detection accuracy is better than other mainstream target detection models,which provide a new method for detecting defects in bamboo slices.

关 键 词：缺陷检测 深度学习 空间特征采样 协作混合分配训练 计算机视觉 

分 类 号：S781.9[农业科学—木材科学与技术] TP391.41[农业科学—林学]