Due to the existence of domain shift, the detection accuracy of road damage detection will drop significantly when the training images and test images are from different scenes. To address this problem, we propose a domain-invariant feature enhancement domain adaptation method (DIFEDA) based on the You Only Look Once (YOLO) series object detectors. We integrate three domain-invariant feature decoupling (DIFD) modules on the backbone to decouple multi-scale domain-invariant features through two-stage adversarial learning. The decoupled features are fed back to the backbone to realize domain-invariant feature enhancement. We construct a spatial and frequency domain perception (SPFD) module in the DIFD to decouple local and global domain-invariant features from the spatial level and frequency level, respectively. We also design a region segmentation decoder (RSD) to make the DIFD pay more attention to the domain-invariant feature extraction in the damaged region, thereby suppressing the interference of background information. We apply DIFEDA to YOLOv8, YOLOv9, YOLO11, and YOLOv12, and conduct extensive experiments in three cross-scenes. The experimental results show that DIFEDA can significantly improve the performance of all baselines, with up to 8.5% and 9.8% improvement [email protected] andF1, respectively, proving the effectiveness and generalization of our method.

Computer vision techniques are the most commonly used methods in automatic road damage detection. However, road damage detection in front-view images is a challenging task due to the complex background. In order to localize the damages more accurately in complex background images, we propose a two-stage region guidance method. In the first-stage region guidance network, a region segmentation task is added to the object detection network using path and weight sharing. The segmented region information is aggregated with the features in the backbone through spatial attention to guide the backbone to enhance the feature extraction in damaged regions. In the second-stage region guidance network, the region-enhanced features in the backbone are efficiently fused with the outputs of neck through the region feature fusion modules to further guide the model for precise localization. We use YOLOv8 as the baseline and propose a road damage detection network based on region guidance, called RDD-RGNet. Experiments on the RDD2020 and CNRDD datasets show that the mAP @0.5 can be improved by up to 2.1%, the F 1- Score can be improved by up to 2.0%, and the Precision can be improved by up to 2.8% compared with the baseline. Due to the path sharing approach, our model can achieve a good balance between the detection accuracy and the number of parameters. The code is available at https://github.com/rodgers-li/RDD-RGNet .

Currently, most neural networks for automatic road damage detection are trained using labeled normal weather datasets, and the detection performance decreases dramatically in hazy and rainy weather, while labeling adverse weather road damage datasets is a rather difficult task. Therefore, we propose a regional feature enhancement domain adaptation method (RFEDA), which employs a two-stage strategy of inter-domain adaptation and intra-domain adaptation to alleviate the domain shift between normal weather and hazy and rainy weather road damage images. Specifically, RFEDA uses a regional feature enhancement module (RFEM) to segment the damage region. The semantic segmentation task facilitates the object detection task to locate the damage object more accurately through path sharing, thus suppressing the interference of background noise. Meanwhile, the segmented region features are utilized to construct instance-level features for instance-level feature alignment. In addition, the multi-scale features extracted by backbone are used for image-level feature alignment. In the intra-domain adaptation stage, the model after inter-domain adaptation is used to generate high-confidence pseudo-labels for the training set of the target domain, and the pseudo-labels are used for self-training of the model and thus for fine-tuning the model. Extensive experiments on the CNRDD and Japan-RDD datasets in hazy and rainy weather demonstrate the effectiveness of our method. Compared with no domain adaptation, RFEDA can improve [email protected] by up to 16.9% and 7.4% in hazy and rainy weather, respectively.

Crack detection is a fundamental effort to ensure road driving safety, aiming to detect potential safety hazards and avoid serious accidents. However, cracks can not be extracted completely and accurately due to problems such as low contrast, high noise, and complex topology of pavement cracks. To address these issues, we propose a directional connectivity feature enhancement network for pavement crack detection. In this network, we build multi-directional enhanced convolution to capture the complex topology of cracks, which is more sensitive to long cracks. To leverage both low-level detail information and high-level semantic information in the network, a novel multi-scale fusion attention is constructed to strengthen the mutual guidance of crack information between channels. The directional connectivity is introducted to establish loss module, which enhances the position and direction information between neighbouring pixels and further refines the crack edge features. To validate the effectiveness and accuracy of the proposed method, we experiment on six publicly available crack datasets, DeepCrack, Crack500, CFD, DCD, EdmCrack600 and DCCE. Compared to other networks, our network achieves 2.2% improvement in ODS and 1.8% improvement in MIoU on the DeepCrack dataset, and 1.2% improvement in ODS and 0.9% improvement in MIoU on the Crack500 dataset. Sufficient experimental results show that our network has better crack detection performance.

Despite significant advances in object detection methods for traffic scenes, object detection under adverse weather conditions is still a challenging task. Especially in foggy weather, the presence of fog reduces visibility, thus weakening the feature information of traffic objects in images, and foggy weather occurs frequently. To cope with this problem, we propose an object detection method with physical prior and adaptive weight convolution (AWConv), and evaluate it on datasets such as Foggy Cityscapes and RTTS. We apply gamma correction in the improved defogging algorithm to enhance the key regions in the image, thus improving the separability of the features. Meanwhile, the feature extraction and representation ability of the model is enhanced by an adaptive weighting mechanism, which in turn improves the model detection performance. In addition, we explore the relationship between image quality and detection accuracy and observe that they are not linearly positively correlated. Due to the complexity of traffic objects in foggy weather, we conduct experiments on Foggy Cityscapes (synthetic fog), RTTS (real-world multiple adverse weather), Cityscapes (normal weather), and extended dataset (different fog concentrations) to validate the model's effectiveness, generalization ability, and robustness. Experimental results show that the small model alone improves mean average precision (mAP) by 1.4% with only 24.6 giga floating point operations per second (GFLOPs) on the Foggy Cityscapes dataset, reduces GFLOPs by 3.8 and improves recall (R) by 1.1% on the RTTS dataset.

Road damage detection (RDD) based on front-view images of roads is more in line with practical application scenarios and is suitable for automatic road damage detection systems. The road damage objects in the front-view images have the characteristics of complex background, multi-scale and large aspect ratio, which greatly increase the difficulty of detection. We propose an anchor-free road damage detection model YOLOX-RDD for front-view images. YOLOX is used as the basic network and three optimization strategies are implemented according to the characteristics of road damage objects. The refined switchable atrous convolution (RSAC) is used to adaptively adjust the receptive field according to the size of the object, which can satisfy the requirements of the detection of the damages of multi-scale and large aspect ratio. For unobvious road damage detection in complex background, four feature enhancement attention (FEA) modules are added to the network to extract more salient information and enhance the fusion effect. Two-level adaptive spatial feature fusion (ASFF) is performed by fusing dark2 with the three output feature maps of neck respectively, and the optimal fusion weights are learned through training to further improve the detection capability of multi-scale objects. The experiments on CNRDD, RDD2020 and USRDD datasets demonstrate the effectiveness and high generalization of our method. Compared with the baseline model, the [email protected] can be improved by up to 2.78%, and F1-Score can be improved by up to 2.55%. The FPS can reach up to 90, achieving a balance between detection accuracy and speed.

Abstract For the two‐stage object detector as a faster region‐convolutional neural network (Faster R‐CNN), upgrading the accuracy of object recognition depends on the proposal box, which is generated by the region proposal algorithms. Due to the limitations of the anchor setting of Faster RCNN, the size of the proposal box generated by the region proposal network (RPN) used is large, which would easily cause a great number of overflows in the sliding search. To improve the accuracy of object detection and remit the overflow problem of the anchor box, multi‐scale anchor box and moving overflow anchor box strategies are introduced here. Then, to increase the positive sample range of the foreground, the hierarchical weight cross entropy classification function is set for binary classification in the RPN network. These strategies could improve the accuracy of object detection. The experimental result achieves 76.2% AP on the Pascal VOC 2007(VOC 07) dataset, which is 2.7% higher than the Faster R‐CNN. The result of the Pascal VOC 2012(VOC 12) test, we achieve 75.6% AP , is improved by 2.5% compared with the Faster R‐CNN.

Object detection and recognition is widely used in various fields and have become key technologies in computer vision. The distribution of objects in natural images can be roughly divided into densely stacked objects and scattered objects. Due to the incomplete attributes or features of some objects in densely stacked distributions, some object detectors have missed local area details or low detection accuracy. In this paper, we propose Cross-splitNet, a novel cross-split method for dense object detection and recognition based on candidate box generation. First, an adaptive feature extraction network is constructed. Different datasets are input into convolutional neural networks with various depths, the generalization of the model. Then, the proposed cross-split algorithm is introduced to guide the different deep networks to learn features of images with various densities, according to intermediate object density classification results. Finally, we adopt a feature pyramid network (FPN) subnet to perform multi-scale feature extraction while retaining lower-layer object information and physical characteristics. The model was trained on the COCO 17, VOC 12, and VOC 07 datasets, which contain a large number of object categories. Our network was compared with several two-stage detectors, and the results show that our model achieved an average precision (AP) of 0.819 at 22.9 frames per second (FPS) on the VOC 07+12 dataset. The mean average precision (mAP) of the object detection model with R50+R2-101 backbones on the COCO dataset was increased by 1.9%.

Aiming at the problems of object mutual occlusion and inaccurate confidence in complex traffic environments, we propose an adaptive network object detection algorithm with a perspective box based on density-aware. Firstly, an adaptive convolutional neural network is designed according to the complexity of image density. Res2Net-101 is used as the backbone. For the images with objects concentrated on both sides and occluded each other more, a high-density network is selected to improve the accuracy of object detection, and a low-density network is selected for the images with significant and simple objects. Secondly, in the high-density network, to improve the accuracy of object detection and reduce the missed detection problem caused by occlusion, a perspective box is added to the detection head, and the transparent box is fused with the prediction box to obtain more accurate positioning. Finally, we propose the perspective loss function, which is based on the repulsion loss of focusing smooth function and combined with feature loss and classification loss to form the overall loss of the model. The experimental results show that the model has a good detection effect compared with the state-of-the-art object detection model in complex traffic environments. Without reducing the detection speed, the mAP of PerspectiveNet on the KITTI dataset is 86.2%, which is 3.6% higher than that of VarifocalNet. On the Cityscapes dataset, the mAP is 96.3%, which is 1.7% higher than that of VarifocalNet.

Abstract The Single Shot MultiBox Detector (SSD) is one of the fastest detection algorithms. Although it has achieved good results in detection, it also has the problem of poor detection effect for small targets and occlusion between objects. Here, the authors propose a new target detection method called single‐shot target detection with multi‐scale feature fusion and feature enhancement. Here, the authors introduce multi‐scale feature fusion module, feature enhancement module and efficient channel attention module, and integrate them into the detection module of the original SSD target detection algorithm to improve the ability of network feature extraction. Experimental results on pascal VOC 2007 datasets show that the proposed algorithm works well when the input size is 300 × 300, the detection speed reaches 41.7 frames per second ( FPS ) and the detection accuracy reaches 79.6%, which is 2.4% higher than the original SSD target detection algorithm. When the input size is 512 × 512, the detection accuracy is 81.9%, and the detection speed reaches 36.5 FPS , which is 3.2% higher than the original SSD target detection algorithm. According to the experimental results, our algorithm has a better performance when there are many objects in the image and there is occlusion.

Abstract Concrete surface cracks detection is an important task to ensure the safety of infrastructure. Because of the complexity of background and low contrast of concrete surface, it is difficult to detect the cracks on the concrete surface accurately. To tackle this problem, a multi‐scale dilated convolutional method for concrete surface crack detection is proposed to improve the accuracy of detection. The proposed network is based on the encoder‐decoder structure of U‐Net. Cascade multi‐scale dilated convolutions in the centre of the network is used to get larger receptive field without additional parameters. In the decoder stage, the feature fusion module is used to integrated the multi‐scale and multi‐level side network feature for the final prediction. A large crack dataset is collected as a training set and other three smaller datasets are used for evaluation. Extensive experiments have been conducted on these three crack datasets, which achieves optimal dataset scale ( ODS ) F‐score over 0.84, optimal image scale ( OIS ) F‐score over 0.85 and average precision ( AP ) over 0.86. This algorithm performs better than the current crack detection, edge detection and semantic segmentation methods.

Crack detection for concrete pavement is an important and fundamental task to ensure road safety. However, automatic crack detection is a challenging topic due to the complicated concrete pavement background and the diversity of cracks. Inspired by the latest developments of deep learning in computer vision, we propose a novel crack detection algorithm of concrete pavement based on attention mechanism and multi-features fusion, and make it possible to deal with various cracks in different pavement backgrounds. The proposed network is constructed using the encoder-decoder structure. The architecture of the encoder part is consisted of Res2Net modules with attention mechanism to achieve fast focus of cracks. Cascade and parallel mode dilated convolutions are set as the center part to enlarge the receptive field of feature points without reducing the resolution of the feature maps. The decoder integrates multiple side output feature maps for pavement crack detection in the manner of feature pyramid. We use ODS, OIS and AP to evaluate the performance of our network. To demonstrate the validity and accuracy of the proposed method, we compare it with some existing methods. The experimental results in multiple crack datasets reveal that our method is superior to these methods.

Automatically detecting cracks with uneven strength from a complex background is a valuable and challenging issue. In light of the lost details and the incomplete extracted cracks in the process of crack extraction, we propose a network model with hierarchical feature fusion and connected attention architecture. Firstly, we build the backbone network on the improved DCA-SE-ResNet-50. Then, we propose a method for crack feature fusion, which combines depthwise separable convolution and dilated convolution to recover more crack details. Finally, we design the attention layer which integrates feature map2 with feature map4. The side network incorporates the feature maps of the low convolutional layer and the high convolutional layer at multiple levels to assist in obtaining the final prediction map. Sufficient experimental results demonstrate that our method achieved state-of-the-art performances, best F-score over 0.86, 12 FPS. Besides the effectiveness of our proposed method is verified on CFD, Crack500, and DCD datasets.

Though crack detection is an indispensable task to ensure the safety of various infrastructures, it is often hard to fully and accurately detect cracks due to their complex background noises as well as long and sharp topological features. To solve this problem, we proposed a network that combines Enhanced Convolution and Dynamic Feature Fusion (ECDFFNet) to improve its overall performance in both capturing long-range dependencies and focus on the local details. In this proposed network, the conventional convolution is replaced by an enhanced one, and a strip mixed convolutional module is embedded in the last two stages of its convolution layers, forming the Enhanced Convolution. Multi-scale information can greatly benefit the crack detection task if they are well fused. Current methods assign a fixed weight to the different scale features regardless of the differences between the local details and semantics. We proposed a Dynamic Feature Fusion (DFF) strategy to the adaptive fusion of different scale features. Extensive experiments on color crack image datasets, i.e., Crack500, CFD, and DeepCrack, show that the proposed model achieved ODS (Optimal Dataset Scale) values of 0.788, 0.863 and 0.872, respectively, and maintained a fast speed of 6 FPS on average in the DeepCrack Dataset. Compared with SegNet, HED, RCF, U-Net, U-HDN, DeepCrack, FPHBN, and DeepCrackT, the proposed method made improvements by 11.5, 8.3, 7, 4.9, 0.5, 4.7, 5.6, and 2.6 respectively, in ODS values.

Automatic crack detection is vital for efficient and economical road maintenance. With the explosive development of convolutional neural networks (CNNs), recent crack detection methods are mostly based on CNNs. In this article, we propose a deeply supervised convolutional neural network for crack detection via a novel multiscale convolutional feature fusion module. Within this multiscale feature fusion module, the high-level features are introduced directly into the low-level features at different convolutional stages. Besides, deep supervision provides integrated direct supervision for convolutional feature fusion, which is helpful to improve model convergency and final performance of crack detection. Multiscale convolutional features learned at different convolution stages are fused together to robustly represent cracks, whose geometric structures are complicated and hardly captured by single-scale features. To demonstrate its superiority and generalizability, we evaluate the proposed network on three public crack data sets, respectively. Sufficient experimental results demonstrate that our method outperforms other state-of-the-art crack detection, edge detection, and image segmentation methods in terms of F1-score and mean IU.

Abstract Keypoint‐based object detection is one of the most efficient and speedy methods at present, yet its performance is often worse than the anchor‐based method. Without prior settings in the keypoint‐based method, the huge search space of the keypoints results in the high recall but low precision. In this paper, the wide dual‐path backbone network is introduced as a feature extractor to extract richer original information, which has fewer parameters and better classification performance. Then, the attention fusion module is designed to effectively fuse the dual‐path with the consideration of the respective advantages of the residual‐path and the densely connected path. In order to provide more accurate pixel‐level information for keypoint prediction, the upsample dual‐attention module is proposed to recover the spatial size of the feature map, which integrates multi‐scale of channel‐wise and spatial attention. Compared with other state‐of‐the‐art detectors, this method has achieved accuracy‐efficiency results with fewer parameters, lower FLOPs, and smaller model size. Experimental results show that the proposed wide dual‐path backbone network has achieved 4.98% top1‐error on the CIFAR‐10 classification dataset. On the PASCAL VOC object detection dataset, this model has achieved an accuracy‐efficiency tradeoff result of 78.3% mAP at the speed of 41 FPS.

Abstract Road Tunnels are an important part of the current road transportation infrastructure. As the main form of tunnel lining diseases, cracks are easy to interact with other areas, which seriously affects the safe operation of the tunnel. Due to the similarity of brightness and linearity between surface cracks and lining cracks, the existing crack detection algorithms can not extract cracks accurately and quickly. An algorithm of lining seam crack elimination with rectangular mark is proposed here. First, the line segments in the image are detected by the Line Segment Detector algorithm based on the coarse percolation detection of the crack. Second, the distribution directions are calculated, and cracks from the lining seams are distinguished by the adaptive threshold judgment method. Third, by using the distribution characteristics of pixels, the line segments are extended to form rectangular marks perpendicular to the direction of lining seams. Finally, the marking information is used to remove the lining joints and obtain the real surface cracks of tunnel lining. Experimental results show that the algorithm can quickly and effectively remove any shape distribution of lining seam. The algorithm fills in the blank of concrete tunnel lining surface crack detection technology.

Aiming at the complex computation and time-consuming problem during unordered image stitching, we present a method based on the binary tree and the estimated overlapping areas to stitch images without order in this paper. For image registration, the overlapping areas between input images are estimated, so that the extraction and matching of feature points are only performed in these areas. For image stitching, we build a model of the binary tree to stitch each two matched images without sorting. Compared to traditional methods, our method significantly reduces the computational time of matching irrelevant image pairs and improves the efficiency of image registration and stitching. Moreover, the stitching model of the binary tree proposed in this paper further reduces the distortion of the panorama. Experimental results show that the number of extracted feature points in the estimated overlapping area is approximately 0.3∼0.6 times of that in the entire image by using the same method, which greatly reduces the computational time of feature extraction and matching. Compared to the exhaustive image matching method, our approach only takes about 1/3 of the time to find all matching images.

Image processing basis crack detection methods can overcome the deficiency of traditional manual and instrument basis crack detection methods, and can provide an important guiding for damage evaluation. However, for tunnel lining surface, cracks and linear seams have great similarities in both intensity value and texture features, it is difficult for existing crack extraction methods to obtain accurate crack segmentation results. In this paper, we proposed a novel algorithm to adaptively eliminate linear seams in tunnel lining crack images. By analyzing characteristics of linear seams and cracks, the idea of binning is used to classify those detected line edges with disordered directions into multiple angle subintervals. Then, by calculating the ratio of statistical specific pixels (SSP ratio) on the expanded line edge, the length and binning information are used to select linear seam edges with the use of adaptive expansion algorithm. Finally, the fragmented segmentations of linear seams are adopted so that cracks and linear seams can be clearly separated, and linear seams can be removed. The experimental results demonstrate the superior precision and efficiency of our method compared with existing methods.

Traditional image stitch methods based on feature point detection require long registration time for high resolution images. In the SIFT (scale invariant feature transform) algorithm, it builds difference‐of‐Gaussian by a linear Gauss expansion filter to obtain the feature. SIFT has poor real time. In this study, a novel image registration method based on image block is proposed to make a rough match for the blocked image and fine match in the most similar blocks by taking advantage of the FAST (features from accelerated segment test) algorithm which runs faster. This way can avoid spending a lot of time in the ineffective area, and enhance the precision and efficiency of the feature point. The accumulative error exists in the process of image stitching, so the image stitched by multiple images has wavelike effects, tilt, or distortion. The camera calibration method is utilised to eliminate the tilt and distortion of the image. The algorithm combining the optimal seam and multi‐resolution fusion is adopted to fuse the stitched image and realise seamless stitch of multiple images in order to achieve a seamless image of high resolution. Simulation experimental results show that the stitching method could realise seamless stitching and straightening of multiple images.