Multi-Layer Feature Extraction with Deformable Convolution for Fabric Defect Detection

In the textile industry,the presence of defects on the surface of fabric is an essential factor in determining fabric quality.Therefore,identifying fabric defects forms a crucial part of the fabric production process.Traditional fabric defect detection algorithms can only detect specific materials and specific fabric defect types;in addition,their detection efficiency is low,and their detection results are relatively poor.Deep learning-based methods have many advantages in the field of fabric defect detection,however,such methods are less effective in identifying multiscale fabric defects and defects with complex shapes.Therefore,we propose an effective algorithm,namely multilayer feature extraction combined with deformable convolution(MFDC),for fabric defect detection.In MFDC,multi-layer feature extraction is used to fuse the underlying location features with high-level classification features through a horizontally connected top-down architecture to improve the detection of multi-scale fabric defects.On this basis,a deformable convolution is added to solve the problem of the algorithm’s weak detection ability of irregularly shaped fabric defects.In this approach,Roi Align and Cascade-RCNN are integrated to enhance the adaptability of the algorithm in materials with complex patterned backgrounds.The experimental results show that the MFDC algorithm can achieve good detection results for both multi-scale fabric defects and defects with complex shapes,at the expense of a small increase in detection time.

A Remote Sensing Image Semantic Segmentation Method by Combining Deformable Convolution with Conditional Random Fields

Currently,deep convolutional neural networks have made great progress in the field of semantic segmentation.Because of the fixed convolution kernel geometry,standard convolution neural networks have been limited the ability to simulate geometric transformations.Therefore,a deformable convolution is introduced to enhance the adaptability of convolutional networks to spatial transformation.Considering that the deep convolutional neural networks cannot adequately segment the local objects at the output layer due to using the pooling layers in neural network architecture.To overcome this shortcoming,the rough prediction segmentation results of the neural network output layer will be processed by fully connected conditional random fields to improve the ability of image segmentation.The proposed method can easily be trained by end-to-end using standard backpropagation algorithms.Finally,the proposed method is tested on the ISPRS dataset.The results show that the proposed method can effectively overcome the influence of the complex structure of the segmentation object and obtain state-of-the-art accuracy on the ISPRS Vaihingen 2D semantic labeling dataset.

A Deformable Network with Attention Mechanism for Retinal Vessel Segmentation

The intensive application of deep learning in medical image processing has facilitated the advancement of automatic retinal vessel segmentation research.To overcome the limitation that traditional U-shaped vessel segmentation networks fail to extract features in fundus image sufficiently,we propose a novel network(DSeU-net)based on deformable convolution and squeeze excitation residual module.The deformable convolution is utilized to dynamically adjust the receptive field for the feature extraction of retinal vessel.And the squeeze excitation residual module is used to scale the weights of the low-level features so that the network learns the complex relationships of the different feature layers efficiently.We validate the DSeU-net on three public retinal vessel segmentation datasets including DRIVE,CHASEDB1,and STARE,and the experimental results demonstrate the satisfactory segmentation performance of the network.

Optical Flow with Learning Feature for Deformable Medical Image Registration

Deformable medical image registration plays a vital role in medical image applications,such as placing different temporal images at the same time point or different modality images into the same coordinate system.Various strategies have been developed to satisfy the increasing needs of deformable medical image registration.One popular registration method is estimating the displacement field by computing the optical flow between two images.The motion field(flow field)is computed based on either gray-value or handcrafted descriptors such as the scale-invariant feature transform(SIFT).These methods assume that illumination is constant between images.However,medical images may not always satisfy this assumption.In this study,we propose a metric learning-based motion estimation method called Siamese Flow for deformable medical image registration.We train metric learners using a Siamese network,which produces an image patch descriptor that guarantees a smaller feature distance in two similar anatomical structures and a larger feature distance in two dissimilar anatomical structures.In the proposed registration framework,the flow field is computed based on such features and is close to the real deformation field due to the excellent feature representation ability of the Siamese network.Experimental results demonstrate that the proposed method outperforms the Demons,SIFT Flow,Elastix,and VoxelMorph networks regarding registration accuracy and robustness,particularly with large deformations.

DSD-MatchingNet:Deformable sparse-to-dense feature matching for learning accurate correspondences

Background Exploring correspondences across multiview images is the basis of various computer vision tasks.However,most existing methods have limited accuracy under challenging conditions.Method To learn more robust and accurate correspondences,we propose DSD-MatchingNet for local feature matching in this study.First,we develop a deformable feature extraction module to obtain multilevel feature maps,which harvest contextual information from dynamic receptive fields.The dynamic receptive fields provided by the deformable convolution network ensure that our method obtains dense and robust correspondence.Second,we utilize sparse-to-dense matching with symmetry of correspondence to implement accurate pixel-level matching,which enables our method to produce more accurate correspondences.Result Experiments show that our proposed DSD-MatchingNet achieves a better performance on the image matching benchmark,as well as on the visual localization benchmark.Specifically,our method achieved 91.3%mean matching accuracy on the HPatches dataset and 99.3%visual localization recalls on the Aachen Day-Night dataset.

基于改进YOLOv8的交通场景实例分割算法

提出一种基于改进型YOLOv8的实例分割算法(DE-YOLO)。为减少图像中复杂背景的干扰,引入高效多尺度注意力机制,跨维交互使各特征组内空间语义特征平均分布。在主干网络部分,使用可变形卷积DCNv2结合C2f卷积层,突破原始卷积限制,提升可变性。为减小有害梯度并提升检测器精度,采用动态非单调聚焦机制Wise-交并比(WIoU)替代联合完全交并(CIoU)损失函数进行质量评估,优化检测框定位,提升分割精度。同时,通过开启Mixup数据增强处理,充实数据集,丰富训练特征,提升模型学习能力。实验结果表明,DE-YOLO在城市景观数据集Cityscapes中的掩模平均精度均值(mAPmask)较基准模型YOLOv8n-seg提高了2.0百分点,IoU阈值为0.5时的平均精度提升了3.2百分点,所提算法在提升精度的同时,保持了优良的检测速度和较少的参数量,模型参数量较同类模型低2.2~31.3百分点。

基于改进YOLOv8n的雨天场景中飞机铆钉检测方法

为解决雨天场景中飞机表面附着与铆钉大小、形状相似的水滴而导致机务工程师在绕机检查过程中易出现铆钉误检的问题,提出1种基于改进YOLOv8n的飞机铆钉小目标检测方法。首先,改进C2f层,融入动态蛇形卷积,以捕捉复杂多变的全局形态特征;其次,在主干网络中嵌入可变形注意力机制,自适应调整对不同区域的关注度;然后,增加1个160×160的小目标检测层,提高小目标的检测能力;最后,使用斯库拉交并比(SIoU)边界框损失函数,提升模型训练速度和推理准确性,基于自建的飞机铆钉和雨滴数据集进行消融实验和对比实验。研究结果表明:本文所提算法在雨天场景下的铆钉检测精确度、召回率、mAP值分别较YOLOv8n提升7.4,4.0,7.8百分点,较其他主流算法也有显著提升。研究结果可为特殊天气下的飞机铆钉检测提供理论基础。

可变形卷积网络的解释性研究及其在蝴蝶物种识别模型中的应用

近年来,可变形卷积网络(DCN)广泛运用于图像识别和分类等领域,然而对该模型的可解释性研究较为有限,它的适用性缺乏充分理论支持。针对上述问题,提出DCN的解释性研究及其在蝴蝶物种识别模型中的应用。首先,引入可变形卷积对VGG16、ResNet50和DenseNet121(Dense Convolutional Network121)分类模型进行改进;其次,采用反卷积和类激活映射(CAM)等可视化手段来对比可变形卷积和标准卷积在特征提取能力上的差异,且通过消融实验结果表明可变形卷积在神经网络的较低层且不连续使用时效果更佳;再次,提出显著性移除(SR)并对CAM的性能和激活特征重要性进行统一评价,同时通过设置不同的移除阈值等多个角度,提高评价的客观性;最后,基于评价结果更高的FullGrad(Full Gradient-weighted)解释模型识别的判断依据。实验结果显示,在Archive_80数据集上,所提出的D_v2-DenseNet121的准确率达到97.03%,相较于DenseNet121分类模型提高了2.82个百分点。可见,可变形卷积的引入赋予了神经网络模型不变性特征提取能力,并提高了分类模型的准确率。

基于金字塔可形变卷积的多分支视频超分模型

为利用帧间的空时相关性特点,提升红外视频超分辨率重建效果,提出了一种改进BasicVSR的超分辨率重建方法。首先,使用金字塔可形变对齐代替BasicVSR中使用的光流法进行帧对齐,将参考帧和相邻帧当作输入,使用可形变卷积对帧间的偏移量进行测量,使不同帧进行信息上的叠加,最大限度得到图像中的细节特征。其次,在上采样时,将参考图像与经过融合后的图像进行级联,通过浅层特征与深层特征的融合,增强特征表达能力。文章设计的模型具有轻量、运行时间短、重建图像主观视觉效果好等优点,且峰值信噪比(PSNR)与结构相似度(SSIM)以及模型运行时间等客观评价指标得到了改进。本文所提模型EbasicVSR比相关模型运行时间平均提升了19 s,信噪比提升了0.14 dB以上,结构相似度提升了2.9%以上,实验结果表明,相比于原BasicVSR模型,本文提出的模型取得了更好的重建效果。

融合可变形卷积的轻量级路面病害检测算法

针对现有的路面病害检测算法在复杂环境下检测精度低、模型复杂度高的问题,在YOLOv5基础上,提出了一种融合可变形卷积的轻量级路面病害检测算法(lightweight deformable convolution YOLOv5,LDC-YOLOv5)。首先,针对真实路面病害复杂不规整的特点,使用可变形卷积(Deformable Conv)和深度卷积(Depthwise Conv),设计了一种轻量级特征提取模块,代替原网络主干部分的C3模块,使卷积核聚焦在无规则裂缝病害上,增强病害特征提取能力。其次,针对特征融合阶段出现算法复杂度过高的问题,使用轻量级卷积GhostConv,构建一种轻量级特征融合模块,代替原网络颈部网络部分的C3模块,降低网络参数和复杂度;为避免真实路面出现光照不均,出现阴影遮挡路面病害目标而造成的病害漏检的情况,在主干网络部分,引入轻量级注意力机制TripletAttention,增强算法对病害信息上下文之间的理解能力。最后在IEEE公开数据集RDD2022和Kaggle公开数据集Road Damage上进行测试,实验结果表明,与YOLOv5s相比,mAP50在两个数据集上分别提升了1.4%和4.2%,且模型参数量仅为YOLOv5s的67.6%。

基于改进ResNet50的岩心图像分类研究

岩心岩性是反映地质条件的重要指标,传统的岩性鉴定通常依赖于人工目视检查,既费时又对专业水平要求高。近年来,卷积神经网络技术的迅速进步,为岩心图像的自动化预测开辟了一条新的途径。本文提出了一种基于改进的ResNet50网络结构的岩心图像分类算法,通过引入ECA(Efficient Channel Attention)注意力机制和PSA (Pyramid Scene Attention)注意力机制,改善了网络对岩心图像丰富地质信息的提取和理解能力,对提高岩性分类的准确性和客观性起到了重要作用;引入可变形卷积(DCNv2),使模型能够自动适应图像特征不规则性和形状变化,显著提升了对岩心结构复杂性的识别能力;使用迁移学习方法,提高了模型的泛化能力和训练效率。实验结果表明,改进的ResNet50网络模型在岩心图像分类任务上表现优异,相较于其他主流卷积网络,平均准确率明显提升,较基线网络ResNet50提高了2.33%的准确率,也有效地提高了对复杂岩心结构的识别精度与鲁棒性。

改进YOLOv8n的无人机航拍图像检测算法

针对无人机航拍图像中目标小、尺度变化大和背景干扰等因素导致检测精度低、定位不准确的问题,提出一种改进YOLOv8n的无人机航拍图像目标检测算法。首先改进C2f模块,利用可变形卷积(DCN)替换其Bottleneck中的卷积以适应航拍图像中物体的形变和尺度变化,同时,在主干网络引入LSK注意力机制,实现动态调整空间感受野,从而在特征提取阶段更灵活地适应不同目标对背景信息需求的差异;然后改进颈部网络,增加一个较浅的检测层并移除大目标检测层,使网络能更有效地捕获小目标的特征以提升检测精度;最后引入WIoU损失函数,使模型更加关注低质量样本,得到更高的检测精度。在VisDrone2019数据集上进行对比实验和消融实验,mAP_(50)值较基线算法模型提升了5.2个百分点,参数量减少了20%,检测速度(FPS)达到87帧/s,能够满足实时性的检测需求。与主流算法进行对比实验,所提算法表现优于目前的主流算法。在DOTA数据集上进行泛化实验,mAP_(50)值提升了1.7个百分点,证明所提算法具有通用性。

一种针对SAR图像的舰船目标检测算法

由于环境复杂、舰船目标散焦和尺度的多样性,基于SAR图像的舰船目标检测仍然存在一些问题。提出了一种针对SAR图像的舰船目标检测算法。首先,基于可变形卷积构建舰船目标特征细化模块,提高对大长宽比姿态的舰船目标的特征提取能力;其次,在主干网络末尾引入了舰船空间金字塔聚合结构,增强对舰船目标的全局特征提取能力;最后,设计了尺度扩展特征金字塔网络,增强舰船浅层和深层特征信息的交互,提高对多尺度舰船目标的检测能力。实验结果表明,所提算法在HRSID数据集上的mAP达到了93.72%,F1分数达到了89.70%,优于所有比较算法,具有良好的检测效果。

An intelligent singular value diagnostic method for concrete dam deformation monitoring

Extracting implicit anomaly information through deformation monitoring data mining is highly significant to determining dam safety status.As an intelligent singular value diagnostic method for concrete dam deformation monitoring, shallow neural network models result in local optima and overfitting, and require manual feature extraction.To obtain an intelligent singular value diagnosis model that can be used for dam safety monitoring, a convolutional neural network (CNN) model that has advantages of deep learning (DL), such as automatic feature extraction, good model fitting, and strong generalizability, was trained in this study.An engineering example shows that the predicted result of the intelligent singular value diagnostic method based on CNN is highly compatible with the confusion matrix, with a precision of 92.41%, receiver operating characteristic (ROC) coordinates of (0.03, 0.97), an area-under-curve (AUC) value of 0.99, and an F1-score of 0.91.Moreover, the performance of the CNN model is better than those of models based on decision tree (DT) and k-nearest neighbor (KNN) methods.Therefore, the intelligent singular value diagnostic method based on CNN is simple to operate, highly intelligent, and highly reliable, and it has a high potential for application in engineering.

LDNet:structure-focused lane detection based on line deformation

Lane detection is a fundamental necessary task for autonomous driving.The conventional methods mainly treat lane detection as a pixel-wise segmentation problem,which suffers from the challenge of uncontrollable driving road environments and needs post-processing to abstract the lane parameters.In this work,a series of lines are used to represent traffic lanes and a novel line deformation network(LDNet) is proposed to directly predict the coordinates of lane line points.Inspired by the dynamic behavior of classic snake algorithms,LDNet uses a neural network to iteratively deform an initial lane line to match the lane markings.To capture the long and discontinuous structures of lane lines,1 D convolution in LDNet is used for structured feature learning along the lane lines.Based on LDNet,a two-stage pipeline is developed for lane marking detection:(1) initial lane line proposal to predict a list of lane line candidates,and(2) lane line deformation to obtain the coordinates of lane line points.Experiments show that the proposed approach achieves competitive performances on the TuSimple dataset while being efficient for real-time applications on a GTX 1650 GPU.In particular,the accuracy of LDNet with the annotated starting and ending points is up to99.45%,which indicates the improved initial lane line proposal method can further enhance the performance of LDNet.

注意力可变形卷积网络的木质板材瑕疵识别

为了解决木材缺陷检测中人工成本高、效率低的问题,该文基于可变性卷积网络和注意力机制,提出一种端到端的神经架构模型.首先,可变形卷积网络(Deformable Convolutional Network, DCN)通过将矩形网格转换为变形网格,使模型专注于具有更多有用图像信息的区域.使用可变形卷积网络可以忽略图像特征中不相关的系数,解决了传统卷积在特征中学习更多信息能力有限的问题.然后,将DCN输出馈送到门控循环单元(Gated Recurrent Unit, GRU)层以学习缺陷图像的高级特征.最后,通过关注输入图像的最重要特征,应用注意力机制加强瑕疵区域的高亮度,从而提高模型识别的准确性.使用Matlab平台在4个木质板材缺陷数据集上将该文方法与现有其他方法进行比较分析,该文方法的准确率比其他3种对比方法提高了2.4%~13.2%的维度,灵敏度提高了3.3%~16.6%的维度,特异性提高了4%~21%的维度.实验结果表明,该文方法在检测精度和其他各个性能方面均优于现有方法,最佳准确率为99.2%,证明了该文方法的有效性.

基于CNN-LSTM的大坝变形组合预测模型研究

为了提高大坝变形预测模型精度和泛化能力,建立了一种基于卷积神经网络(Convolutional neural networks,CNN)与深度学习长短期记忆(Long short-term memory,LSTM)神经网络的组合预测模型CNN-LSTM。该模型先利用CNN提取大坝变形监测时间序列的特征,再利用LSTM生成特征描述,该模型精度高、泛化能力强。以柏叶口水库混凝土面板堆石坝为例,经过CNN-LSTM模型计算,将模型变形预测值与原型监测资料进行对比,再与LSTM模型及CNN模型的预测结果进行对比。结果表明,CNN-LSTM模型预测值最接近监测资料实测结果。

基于改进的Transformer细粒度图像识别算法研究

针对细粒度图像识别存在类间差异小、难以区分等问题,本文通过提升网络对图像细节特征的表达能力,来改善这一问题。为此,设计了一种基于改进的Transformer细粒度识别算法。首先,可变形卷积令牌嵌入通过自适应调整采样点的位置,来改变卷积操作范围及其卷积核的形状,从而增强网络模型对空间信息的感知能力,以获取更为精准的空间信息;其次,高效相关通道注意力机制通过对通道的自动选择,将通道注意力的计算从通道相邻转换成语义相似,来捕获语义相似的通道信息。而精准的空间信息和语义相似的通道信息将有效提升网络模型局部特征感知能力。实验结果表明,与基线算法相比,本文方法在CUB-200-2011、StanfordCars和StanfordDogs三个数据集上的识别结果分别提升了1.5%、2.4%、1.5%。结果表明,本文提出的方法通过提升细粒度图像细节特征的表达能力,从而有效提高了细粒度图像识别的有效性。

基于TCN和迁移学习的混凝土坝变形预测方法

混凝土坝变形测点数据丢失或者新增测点测量时间太短都会导致这部分测点的数据量不足,使得变形预测精度受到影响。为了提高这些小数据量测点的变形预测精度,提出了将时域卷积网络(TCN)与迁移学习相结合的变形预测方法。以数据量充足的测点为源域,以缺少数据的测点为目标域,将在源域上训练好的TCN模型的结构和参数迁移到目标域模型中,固定其中的冻结层参数,利用目标域中的数据对目标域模型可调层参数进行调整。同时,采用动态时间规整选择与目标域数据序列相似度最高的监测数据作为最佳源域数据,提升迁移学习效果。工程实例分析表明:迁移学习后的目标域模型的均方根误差和平均绝对误差与利用足量数据训练的TCN模型的预测误差相比,差异仅分别为1.73%和8.09%,小数据量情况下TCN预测模型的精度得到了提高。

基于改进YOLO v5算法的道路小目标检测

为解决交通道路小目标检测难度大、精度低,容易出现错检漏检的问题,提出一种基于YOLO v5(you only look once v5)算法的多尺度特征融合目标检测改进算法。首先,增加小目标检测头用于适应小目标尺寸,缓解漏检情况。然后,引入可变形卷积网络v2(deformable convolutional networks V2,DCN V2)提高模型对运动中小目标的学习能力;同时,增加上下文增强模块,提升对远距离小目标的识别能力。最后,在替换损失函数、提高边界框定位精度的同时,使用空间金字塔池化和上下文空间金字塔卷积分组模块,提高网络的感受野和特征表达能力。实验结果表明,所提算法在KITTI数据集小目标类别上平均识别精度达到了95.2%,相较于原始YOLO v5,算法总体平均识别精度提升了2.7%,对小目标的检测效果更佳,平均识别精度提升了3.1%,证明所提算法在道路小目标检测方面的有效性。