Bounding Box Research Articles

Tiny Object Tracking: A Large-Scale Dataset and a Baseline.

Tiny objects, frequently appearing in practical applications, have weak appearance and features, and receive increasing interests in many vision tasks, such as object detection and segmentation. To promote the research and development of tiny object tracking, we create a large-scale video dataset, which contains 434 sequences with a total of more than 217K frames. Each frame is carefully annotated with a high-quality bounding box. In data creation, we take 12 challenge attributes into account to cover a broad range of viewpoints and scene complexities, and annotate these attributes for facilitating the attribute-based performance analysis. To provide a strong baseline in tiny object tracking, we propose a novel multilevel knowledge distillation network (MKDNet), which pursues three-level knowledge distillations in a unified framework to effectively enhance the feature representation, discrimination, and localization abilities in tracking tiny objects. Extensive experiments are performed on the proposed dataset, and the results prove the superiority and effectiveness of MKDNet compared with state-of-the-art methods. The dataset, the algorithm code, and the evaluation code are available at https://github.com/mmic-lcl/Datasets-and-benchmark-code.

Spatio-temporal action detector with cross-stream knowledge transfer

Abstract Relying on human labor for uninterrupted monitoring through surveillance systems is highly inefficient and prone to oversight, making intelligent surveillance have great application prospects and value. To implement this technology, spatio-temporal action detectors are typically used, which are neural networks capable of locating action instances simultaneously in both time and space dimensions. To address the issue of high computational demand commonly associated with 3D backbone architectures, we propose a Dynamic Cross-Stream MovingCenter Detector (DCMOC) based on a 2D backbone network, an improved version of the MOC detector. In the DCMOC detector, the movement branch and the bounding box branch correspond to the motion and appearance features of action instances, respectively. Through a designed cross-stream knowledge transfer module, feature knowledge is transferred between the two branches to enhance feature representation. To compare with existing spatio-temporal action detectors, experiments were conducted on two datasets, JHMDB and UCF101-24. The experiments demonstrate that the DCMOC detector achieves competitive performance with significantly lower computational requirements than existing 3D architecture models and achieves optimal performance on the video-map metric at high thresholds.

A Three-Step Computer Vision-Based Framework for Concrete Crack Detection and Dimensions Identification

Crack detection is significant to building repair and maintenance; however, conventional inspection is a labor-intensive and time-consuming process for field engineers. This paper proposes a three-step computer vision-based framework to quickly recognize concrete cracks and automatically identify their length, maximum width, and area in damage images. In step one, a region-based convolutional neural network (YOLOv8) is applied to train the crack localizing model. In step two, Gaussian filtering, Canny, and FindContours are integrated to extract the reference contour (a pre-designed seal) to obtain the conversion scale between pixels and millimeter-wise sizes. In step three, the recognized crack bounding box is cropped, and the ApproxPolyDP function and Hough transform are performed to quantify crack dimensions based on the conversion ratio. The developed framework was validated on a dataset of 4630 crack images, and the model training took 150 epochs. Results show that the average crack detection accuracy reaches 95.7%, and the precision of quantified dimensions is over 90%, while the error increases as the crack size grows smaller (increasing to 8% when the crack width is within 1 mm). The proposed method can help engineers to efficiently achieve crack information at building inspection sites, while the reference frame must be pre-marked near the crack, which may limit the scope of application scenarios. In addition, the robustness and accuracy of the developed image processing techniques-based crack quantification algorithm need to be further improved to meet the requirements in real cases when the crack is located within a complex background.

SD-YOLO-AWDNet: A hybrid approach for smart object detection in challenging weather for self-driving cars

Several deep learning algorithms are currently focused on object detection in adverse weather scenarios for autonomous driving systems. However, these algorithms face challenges in real-time scenarios, leading to a reduction in detection accuracy. To tackle these issues, this paper introduces a lightweight object detection model named Self Driving Cars You Only Look Once Adverse Weather Detection Network (SD-YOLO-AWDNet), derived from enhancements to the YOLOv5 algorithm. The model incorporates four progressive improvement levels within the YOLOv5 framework. This includes integrating C3Ghost and GhostConv modules in the backbone to enhance detection speed by reducing computational overhead during feature extraction. To address potential accuracy issues arising from these modules, Depthwise-Separable Dilated Convolutions (DSDC) are introduced, striking a balance between accuracy and parameter reduction. The model further incorporates a Coordinate Attention (CA) module in the GhostBottleneck to enhance feature extraction and eliminate unnecessary features, improving precision in object detection. Additionally, a novel “Focal Distribution Loss” replaces CIoU Loss, accelerating bounding box regression and loss reduction. Test dataset experiments demonstrate that SD-YOLO-AWDNet outperforms YOLOv5 with a 54% decrease in FLOPs, a 52.53% decrease in model parameters, a 2.24% increase in mAP, and a threefold improvement in detection speed.

Innovative healthcare solutions: robust hand gesture recognition of daily life routines using 1D CNN.

Hand gestures are an effective communication tool that may convey a wealth of information in a variety of sectors, including medical and education. E-learning has grown significantly in the last several years and is now an essential resource for many businesses. Still, there has not been much research conducted on the use of hand gestures in e-learning. Similar to this, gestures are frequently used by medical professionals to help with diagnosis and treatment. We aim to improve the way instructors, students, and medical professionals receive information by introducing a dynamic method for hand gesture monitoring and recognition. Six modules make up our approach: video-to-frame conversion, preprocessing for quality enhancement, hand skeleton mapping with single shot multibox detector (SSMD) tracking, hand detection using background modeling and convolutional neural network (CNN) bounding box technique, feature extraction using point-based and full-hand coverage techniques, and optimization using a population-based incremental learning algorithm. Next, a 1D CNN classifier is used to identify hand motions. After a lot of trial and error, we were able to obtain a hand tracking accuracy of 83.71% and 85.71% over the Indian Sign Language and WLASL datasets, respectively. Our findings show how well our method works to recognize hand motions. Teachers, students, and medical professionals can all efficiently transmit and comprehend information by utilizing our suggested system. The obtained accuracy rates highlight how our method might improve communication and make information exchange easier in various domains.

An Automated Diagnosis Method for Lung Cancer Target Detection and Subtype Classification-Based CT Scans.

When dealing with small targets in lung cancer detection, the YOLO V8 algorithm may encounter false positives and misses. To address this issue, this study proposes an enhanced YOLO V8 detection model. The model integrates a large separable kernel attention mechanism into the C2f module to expand the information retrieval range, strengthens the extraction of lung cancer features in the Backbone section, and achieves effective interaction between multi-scale features in the Neck section, thereby enhancing feature representation and robustness. Additionally, depth-wise convolution and Coordinate Attention mechanisms are embedded in the Fast Spatial Pyramid Pooling module to reduce feature loss and improve detection accuracy. This study introduces a Minimum Point Distance-based IOU loss to enhance correlation between predicted and ground truth bounding boxes, improving adaptability and accuracy in small target detection. Experimental validation demonstrates that the improved network outperforms other mainstream detection networks in terms of average precision values and surpasses other classification networks in terms of accuracy. These findings validate the outstanding performance of the enhanced model in the localization and recognition aspects of lung cancer auxiliary diagnosis.

RA-YOLOv8: An Improved YOLOv8 Seal Text Detection Method

To detect text from electronic seals that have significant background interference, blurring, text overlapping, and curving, an improved YOLOv8 model named RA-YOLOv8 was developed. The model is primarily based on YOLOv8, with optimized structures in its backbone and neck. The receptive-field attention and efficient multi-scale attention (RFEMA) module is introduced in the backbone. The model’s ability to extract and integrate local and global features is enhanced by combining the attention on the receptive-field spatial feature of the receptive-field attention and coordinate attention (RFCA) module and the cross-spatial learning of the efficient multi-scale attention (EMA) module. The Alterable Kernel Convolution (AKConv) module is incorporated in the neck, enhancing the model’s detection accuracy of curved text by dynamically adjusting the sampling position. Furthermore, to boost the model’s detection performance, the original loss function is replaced with the bounding box regression loss function of Minimum Point Distance Intersection over Union (MPDIoU). Experimental results demonstrate that RA-YOLOv8 surpasses YOLOv8 in terms of precision, recall, and F1 value, with improvements of 0.4%, 1.6%, and 1.03%, respectively, validating its effectiveness and utility in seal text detection.

COTTON-YOLO: Enhancing Cotton Boll Detection and Counting in Complex Environmental Conditions Using an Advanced YOLO Model

This study aims to enhance the detection accuracy and efficiency of cotton bolls in complex natural environments. Addressing the limitations of traditional methods, we developed an automated detection system based on computer vision, designed to optimize performance under variable lighting and weather conditions. We introduced COTTON-YOLO, an improved model based on YOLOv8n, incorporating specific algorithmic optimizations and data augmentation techniques. Key innovations include the C2F-CBAM module to boost feature recognition capabilities, the Gold-YOLO neck structure for enhanced information flow and feature integration, and the WIoU loss function to improve bounding box precision. These advancements significantly enhance the model’s environmental adaptability and detection precision. Comparative experiments with the baseline YOLOv8 model demonstrated substantial performance improvements with COTTON-YOLO, particularly a 10.3% increase in the AP50 metric, validating its superiority in accuracy. Additionally, COTTON-YOLO showed efficient real-time processing capabilities and a low false detection rate in field tests. The model’s performance in static and dynamic counting scenarios was assessed, showing high accuracy in static cotton boll counting and effective tracking of cotton bolls in video sequences using the ByteTrack algorithm, maintaining low false detections and ID switch rates even in complex backgrounds.

Point Cloud Measurement of Rubber Tread Dimension Based on RGB-Depth Camera

To achieve an accurate measurement of tread size after fixed-length cutting, this paper proposes a point-cloud-based tread size measurement method. Firstly, a mathematical model of corner points and a reprojection error is established, and the optimal solution of the number of corner points is determined by the non-dominated sorting genetic algorithm II (NSGA-II), which reduces the reprojection error of the RGB-D camera. Secondly, to address the problem of the low accuracy of the traditional pixel metric ratio measurement method, the random sampling consensus point cloud segmentation algorithm (RANSAC) and the oriented bounding box (OBB) collision detection algorithm are introduced to complete the accurate detection of the tread size. By comparing the absolute error and relative error data of several groups of experiments, the accuracy of the detection method in this paper reaches 1 mm, and the measurement deviation is between 0.14% and 2.67%, which is in line with the highest accuracy standard of the national standard. In summary, the RGB-D visual inspection method constructed in this paper has the characteristics of low cost and high inspection accuracy, which is a potential solution to enhance the pickup guidance of tread size measurement.

Weakly Supervised Object Detection in Chest X-Rays with Differentiable ROI Proposal Networks and Soft ROI Pooling.

Weakly supervised object detection (WSup-OD) increases the usefulness and interpretability of image classification algorithms without requiring additional supervision. The successes of multiple instance learning in this task for natural images, however, do not translate well to medical images due to the very different characteristics of their objects (i.e. pathologies). In this work, we propose Weakly Supervised ROI Proposal Networks (WSRPN), a new method for generating bounding box proposals on the fly using a specialized region of interest-attention (ROI-attention) module. WSRPN integrates well with classic backbone-head classification algorithms and is end-to-end trainable with only image-label supervision. We experimentally demonstrate that our new method outperforms existing methods in the challenging task of disease localization in chest X-ray images. Code: https://anonymous.4open.science/r/WSRPN-DCA1.

Automatic anal sphincter integrity detection from ultrasound images via convolutional neural networks.

The anal sphincter complex comprises the anal sphincter and the U-shaped deep and superficial puborectalis muscle. As an important supporting structure of the posterior pelvic floor, together with its surrounding tissues and muscles, the anal sphincter complex maintains the normal physiological functions of defecation and continence. The plane required for diagnosing anal sphincter injury and the diagnosis of anal sphincter integrity through pelvic floor ultrasound are highly dependent on sonographers' experience. We developed a deep learning (DL) tool for the automatic diagnosis of anal sphincter integrity via pelvic floor ultrasound. A 2D detection network was trained to detect the bounding box of the anal sphincter. The pelvic floor ultrasound image and its corresponding oval mask were input into a 2D classification network to determine the integrity of the anal sphincter. The average precision (AP) and intersection over union (IoU) were used to evaluate the performance of anal sphincter detection. Receiver operating characteristic (ROC) analysis was used to evaluate the performance of the classification model. The Pearson correlation coefficients (r values) of the topmost and bottommost layers detected by the CNN and sonographers were 0.932 and 0.978, respectively. The best DL model yielded the highest area under the curve (AUC) of 0.808 (95% CI: 0.698-0.921) in the test cohort. The results from the CNN agreed well with the diagnostic results of experienced sonographers. We proposed, for the first time, a CNN to obtain the plane required for diagnosing anal sphincter injury on the basis of pelvic floor ultrasound and for preliminarily diagnosing anal sphincter injury.

YOLOv7-Branch: A Jujube Leaf Branch Detection Model for Agricultural Robot.

The intelligent harvesting technology for jujube leaf branches presents a novel avenue for enhancing both the quantity and quality of jujube leaf tea, whereas the precise detection technology for jujube leaf branches emerges as a pivotal factor constraining its development. The precise identification and localization of jujube leaf branches using real-time object detection technology are crucial steps toward achieving intelligent harvesting. When integrated into real-world scenarios, issues such as the background noise introduced by tags, occlusions, and variations in jujube leaf morphology constrain the accuracy of detection and the precision of localization. To address these issues, we describe a jujube leaf branch object detection network based on YOLOv7. First, the Polarized Self-Attention module is embedded into the convolutional layer, and the Gather-Excite module is embedded into the concat layer to incorporate spatial information, thus achieving the suppression of irrelevant information such as background noise. Second, we incorporate implicit knowledge into the Efficient Decoupled Head and replace the original detection head, enhancing the network's capability to extract deep features. Third, to address the issue of imbalanced jujube leaf samples, we employ Focal-EIoU as the bounding box loss function to expedite the regression prediction and enhance the localization accuracy of the model's bounding boxes. Experiments show that the precision of our model is 85%, which is increased by 3.5% compared to that of YOLOv7-tiny. The mAP@0.5 value is 83.7%. Our model's recognition rate, recall and mean average precision are superior to those of other models. Our method could provide technical support for yield estimation in the intelligent management of jujube orchards.

HP-YOLOv8: High-Precision Small Object Detection Algorithm for Remote Sensing Images.

YOLOv8, as an efficient object detection method, can swiftly and precisely identify objects within images. However, traditional algorithms encounter difficulties when detecting small objects in remote sensing images, such as missing information, background noise, and interactions among multiple objects in complex scenes, which may affect performance. To tackle these challenges, we propose an enhanced algorithm optimized for detecting small objects in remote sensing images, named HP-YOLOv8. Firstly, we design the C2f-D-Mixer (C2f-DM) module as a replacement for the original C2f module. This module integrates both local and global information, significantly improving the ability to detect features of small objects. Secondly, we introduce a feature fusion technique based on attention mechanisms, named Bi-Level Routing Attention in Gated Feature Pyramid Network (BGFPN). This technique utilizes an efficient feature aggregation network and reparameterization technology to optimize information interaction between different scale feature maps, and through the Bi-Level Routing Attention (BRA) mechanism, it effectively captures critical feature information of small objects. Finally, we propose the Shape Mean Perpendicular Distance Intersection over Union (SMPDIoU) loss function. The method comprehensively considers the shape and size of detection boxes, enhances the model's focus on the attributes of detection boxes, and provides a more accurate bounding box regression loss calculation method. To demonstrate our approach's efficacy, we conducted comprehensive experiments across the RSOD, NWPU VHR-10, and VisDrone2019 datasets. The experimental results show that the HP-YOLOv8 achieves 95.11%, 93.05%, and 53.49% in the mAP@0.5 metric, and 72.03%, 65.37%, and 38.91% in the more stringent mAP@0.5:0.95 metric, respectively.

Above the roads: Unleashing unmanned aerial vehicles and image processing for traffic analysis

Road traffic surveillance using unmanned aerial vehicles is a practice that can be found especially in the field of intelligent vehicle management, which is still in the early stages of research and application. This paper presents three methods of analyzing traffic data. One method is a conventional one, based on Doppler radar detection and the other two methods analyze images captured by unmanned aerial vehicles, being based on deep learning techniques. After acquiring the images, they went through a complex processing process to eliminate noise and improve the clarity of the image, then the identification of the vehicles was done by recognizing moving objects and highlighting them either through a bounding box or through labelling. The quality of images obtained from unmanned aerial vehicles is similar to the quality of images obtained from fixed surveillance cameras. The comparative analysis of the results obtained through image processing, together with those obtained through a conventional method of traffic analysis, the Doppler radar, highlighted the fact that video detection used in intelligent vehicle management is a method that both researchers and local authorities can rely on the performance of traffic studies or the analysis of traffic incidents and accidents.

Dental cavity analysis, prediction, localization, and quantification using computer vision

Dental health assessment is a critical component of overall well-being, and advancements in computer vision and deep learning have opened new avenues for automating and enhancing this process. In this study, we present a comprehensive approach to dental cavity analysis, spanning localization, quantification, and visualization. Our methodology leveraged a diverse dataset of colored dental images that had been meticulously augmented and annotated. The You Only Look Once model was employed for precise dental cavity localization, providing bounding box predictions. Remarkably, these results were obtained based on images from standard device cameras. Subsequently, we introduced the use of the segment anything model segmentation model, known for its zero-shot generalization capabilities, to focus on the exact areas of dental cavities. This approach enhanced the granularity of our analysis, providing dental professionals with detailed visualizations for precise diagnosis. During the quantification phase, we extracted cavity areas from bounding box coordinates, enabling accurate measurement of cavity sizes. The model achieved a notable mean average precision of 0.732, an accuracy of 0.789, and a recall of 0.701. Moreover, the model converged quickly, with most metrics achieving near-optimal results after 100 iterations. This quantitative data augments traditional diagnosis methods, facilitating more informed treatment decisions.

An Improved Fire and Smoke Detection Method Based on YOLOv8n for Smart Factories.

Factories play a crucial role in economic and social development. However, fire disasters in factories greatly threaten both human lives and properties. Previous studies about fire detection using deep learning mostly focused on wildfire detection and ignored the fires that happened in factories. In addition, lots of studies focus on fire detection, while smoke, the important derivative of a fire disaster, is not detected by such algorithms. To better help smart factories monitor fire disasters, this paper proposes an improved fire and smoke detection method based on YOLOv8n. To ensure the quality of the algorithm and training process, a self-made dataset including more than 5000 images and their corresponding labels is created. Then, nine advanced algorithms are selected and tested on the dataset. YOLOv8n exhibits the best detection results in terms of accuracy and detection speed. ConNeXtV2 is then inserted into the backbone to enhance inter-channel feature competition. RepBlock and SimConv are selected to replace the original Conv and improve computational ability and memory bandwidth. For the loss function, CIoU is replaced by MPDIoU to ensure an efficient and accurate bounding box. Ablation tests show that our improved algorithm achieves better performance in all four metrics reflecting accuracy: precision, recall, F1, and mAP@50. Compared with the original model, whose four metrics are approximately 90%, the modified algorithm achieves above 95%. mAP@50 in particular reaches 95.6%, exhibiting an improvement of approximately 4.5%. Although complexity improves, the requirements of real-time fire and smoke monitoring are satisfied.

Coarse‐to‐fine adjustment for multi‐platform point cloud fusion

AbstractLeveraging multi‐platform laser scanning systems offers a complete solution for 3D modelling of large‐scale urban scenes. However, the spatial inconsistency of point clouds collected by heterogeneous platforms with different viewpoints presents challenges in achieving seamless fusion. To tackle this challenge, this paper proposes a coarse‐to‐fine adjustment for multi‐platform point cloud fusion. First, in the preprocessing stage, the bounding box of each point cloud block is employed to identify potential constraint association. Second, the proposed local optimisation facilitates preliminary pairwise alignment with these potential constraint relationships, and obtaining initial guess for a comprehensive global optimisation. At last, the proposed global optimisation incorporates all the local constraints for tightly coupled optimisation with raw point correspondences. We choose two study areas to conduct experiments. Study area 1 represents a fast road scene with a significant amount of vegetation, while study area 2 represents an urban scene with many buildings. Extensive experimental evaluations indicate the proposed method has increased the accuracy of study area 1 by 50.6% and the accuracy of study area 2 by 44.7%.

Research on the Method and Application of Truck Type Recognition Based on Deep Learning from the Perspective of Unmanned Aerial Vehicles

A lightweight detection model for truck models based on improved YOLOv5s (MobileNetV3-YOLOv5s) is proposed to meet the requirement for real-time detection under the limited embedded device resources carried by drones. Firstly, we use MobileNetV3 to replace the backbone feature extraction network and use deep separable convolution to replace traditional convolution to reduce the model’s parameter count. Secondly, we use DIOU loss as the regression loss of the bounding box to enhance the convergence speed of the model and improve the ability to fit data. Finally, we use the K-means clustering method to reset the prior box. The experimental results show that the mAP value of the improved model is 89.6%, which is 0.2 percentage points lower than before, but the volume is only 3.98MB, which is about half of the original model. The detection speed is also significantly improved compared with before the improvement. Therefore, the lightweight model based on improved YOLOv5s improves detection speed and significantly reduces model volume while ensuring detection accuracy. It enables efficient, real-time recognition of truck models under complex road conditions on embedded devices.

Prototypical contrastive learning based oriented detector for kitchen waste

Automatic detection of kitchen waste enables the identification and quantification of non-degradable materials, such as plastics, metals, and other substances that cannot be easily decomposed. This approach increases the efficiency of the waste disposal process with significant time and cost savings compared to manual sorting. However, detecting non-degradable waste automatically presents a great challenge due to the deformability and directional uncertainty of the objects in kitchen waste images. Object deformation leads to a lack of structural features, making it more difficult to distinguish objects. Direction uncertainty leads to redundant backgrounds when detecting with the horizontal box detector. To address these issues, we propose a kitchen waste detector (KWDet) for the automatic detection of non-degradable waste in kitchen waste. First, the KWDet introduces the oriented bounding box representation and rotated region of interest (RRoI) learner, which enables the oriented detection of non-degradable waste in kitchen waste. Second, the KWDet constructs a prototypical contrastive learning (PCL) based detection head to improve the classification performance by learning discriminative features of different non-degradable waste in kitchen waste. In the PCL based detection head, the prototype-sample contrastive loss is proposed to pull samples and prototypes from the same classes together while pushing samples and prototypes from the different classes apart. After that, we design the inter-prototype contrastive loss to learn the discriminative prototypes of different non-degradable waste. Finally, to evaluate the proposed KWDet, we construct a large kitchen waste detection dataset (KWDD), containing 13,873 kitchen waste images and 213,565 instances collected from Zhengzhou China with 5 classes. To validate the effectiveness of our proposed method, we performed ablation experiments and comparison experiments on the KWDD dataset. The experiments demonstrate that KWDet achieves a state-of-the-art (SOTA) mAP50 of 71.5%, surpassing other object detection methods by more than 1.6% mAP50. Furthermore, our approach’s Macro-F1 score outperforms the SOTA method by 4.0%. Extensive experiments on the KWDD demonstrate that the prototypical contrastive learning-based oriented detector is able to extract more robust features and better characterize the target location, suggesting that the KWDet is a powerful support for intelligent kitchen waste sorting.

Improving Rebar Twist Prediction Exploiting Unified-Channel Attention-Based Image Restoration and Regression Techniques.

Recent research has made significant progress in automated unmanned systems utilizing Artificial Intelligence (AI)-based image processing to optimize the rebar manufacturing process and minimize defects such as twisting during production. Despite various studies, including those employing data augmentation through Generative Adversarial Networks (GANs), the performance of rebar twist prediction has been limited due to image quality degradation caused by environmental noise, such as insufficient image quality and inconsistent lighting conditions in rebar processing environments. To address these challenges, we propose a novel approach for real-time rebar twist prediction in manufacturing processes. Our method involves restoring low-quality grayscale images to high resolution and employing an object detection model to identify and track rebar endpoints. We then apply regression analysis to the coordinates obtained from the bounding boxes to estimate the error rate of the rebar endpoint positions, thereby determining the occurrence of twisting. To achieve this, we first developed a Unified-Channel Attention (UCA) module that is robust to changes in intensity and contrast for grayscale images. The UCA can be integrated into image restoration models to more accurately detect rebar endpoint characteristics in object detection models. Furthermore, we introduce a method for predicting the future positions of rebar endpoints using various linear and non-linear regression models. The predicted positions are used to calculate the error rate in rebar endpoint locations, determined by the distance between the actual and predicted positions, which is then used to classify the presence of rebar twisting. Our experimental results demonstrate that integrating the UCA module with our image restoration model significantly improved existing models in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) metrics. Moreover, employing regression models to predict future rebar endpoint positions enhances the F1 score for twist prediction. As a result, our approach offers a practical solution for rapid defect detection in rebar manufacturing processes.