Predicted Bounding Boxes Research Articles

MACN: A cascade defect detection for complex background based on mixture attention mechanism

AbstractDefect detection in complex background is a critical issue. To address this issue, this paper proposes the mixture attention mechanism cascade network, in which the new channel attention network is linked with the spatial attention network to create an effective mixed attention network that takes advantage of their respective advantages, adaptively suppresses background noise, and highlights defect features. To ensure the efficiency and effectiveness of effective mixed attention network, the new channel attention network splices the output features of the global average pooling layer and the global maximum pooling layer and then sends the spliced features into a shared network, which is a one‐dimensional convolutional network, and uses cross‐channel interaction for fusion. Furthermore, in order to provide more discriminative feature representation, the authors extract the intermediate features of the region proposal network and input them into effective mixed attention network. Finally, the cascade head is used to refine the predicted bounding box to achieve high‐quality defect location. To demonstrate the superiority and usefulness of this method, it is compared to the latest method using widely used PCB and NEU data sets. A large number of trials demonstrate that this strategy outperforms other methods for detecting defects in complicated backgrounds.

FRED: Towards a Full Rotation-Equivariance in Aerial Image Object Detection

Rotation-equivariance is an essential yet challenging property in oriented object detection. While general object detectors naturally leverage robustness to spatial shifts due to the translation-equivariance of the conventional CNNs, achieving rotation-equivariance remains an elusive goal. Current detectors deploy various alignment techniques to derive rotation-invariant features, but still rely on high capacity models and heavy data augmentation with all possible rotations. In this paper, we introduce a Fully Rotation-Equivariant Oriented Object Detector (FRED), whose entire process from the image to the bounding box prediction is strictly equivariant. Specifically, we decouple the invariant task (object classification) and the equivariant task (object localization) to achieve end-to-end equivariance. We represent the bounding box as a set of rotation-equivariant vectors to implement rotation-equivariant localization. Moreover, we utilized these rotation-equivariant vectors as offsets in the deformable convolution, thereby enhancing the existing advantages of spatial adaptation. Leveraging full rotation-equivariance, our FRED demonstrates higher robustness to image-level rotation compared to existing methods. Furthermore, we show that FRED is one step closer to non-axis aligned learning through our experiments. Compared to state-of-the-art methods, our proposed method delivers comparable performance on DOTA-v1.0 and outperforms by 1.5 mAP on DOTA-v1.5, all while significantly reducing the model parameters to 16%.

Geometric information constraint 3D object detection from LiDAR point cloud for autonomous vehicles under adverse weather

3D object detection, as the core of the autonomous vehicle perception module, is essential for efficient transportation and comfortable experiences. However, the challenge of 3D object detection under adverse weather conditions hinders the advancement of autonomous vehicles to higher levels. Hence, achieving accurate 3D object detection under adverse weather conditions is increasingly crucial as it forms the foundation for trajectory planning and driving strategy making in autonomous vehicles, thereby revolutionizing transportation modes for both goods and passengers. Advances in Light Detection and Ranging (LiDAR) technology have facilitated the development of 3D object detection in the past few years. Adverse weather, which inevitably occurs in real-world driving scenarios, could degrade measurement accuracy and point density of LiDAR and lead to particle interference. Detecting accurate 3D bounding boxes from sparse, incomplete point clouds with particle interference is difficult. Therefore, this research presents a novel geometric information constraint network for 3D object detection tasks from LiDAR point clouds under adverse weather (GIC-Net). In this study, we focus on how to incorporate geometric location information and line geometric feature information into the network against adverse weather. Further, we propose a geometric location constrained backbone module (GLC) to reduce rain and snow particle interference and ensure sufficient receptive fields. Then, we design a line geometric feature constraint module (LGFC) to add line constraints of 3D bounding boxes into the training process. Finally, a line loss function is designed, and features from the GLC and LGFC modules are fed into the multi-task detection head for accurate 3D bounding box prediction. Experiments on the Canadian Adverse Driving Conditions (CADC) autonomous vehicle dataset demonstrate the superiority of our method over six other state-of-the-art methods under adverse weather, which is at least 13.32 %, 4.67 %, and 10.44 % mAP higher than the other compared methods in the car, truck, and pedestrian classes respectively. Also, we further verify the better generalization ability of our network compared to other methods.

WildARe-YOLO: A lightweight and efficient wild animal recognition model

For the protection of endangered species and successful wildlife population monitoring, wild animal recognition is essential. While deep learning models like YOLOv5 have shown promise in real-time object recognition, their practical applicability may be constrained by their high processing requirements. In this paper, we suggest a faster and lighter version of YOLOv5s for wild animal recognition. To lower computational costs for model parameters and floating-point operations (FLOPs) for the backbone, our suggested model includes Mobile Bottleneck Block modules and an improved StemBlock. We also use Focal-EIoU as a loss function to gauge the accuracy of the predicted bounding boxes during inference and employ a BiFPN-based neck. We tested our technique on three datasets, including Wild Animal Facing Extinction, Fishmarket, and MS COCO 2017. Additionally, our technique is compared with state-of-the-art deep learning models, and from the baseline model we recorded a 17.65% increase in FPS, 28.55% model parameters reduction, and 50.92% in FLOPs reduction. Furthermore, our model has a faster model loading time, which is critical for deployment in remote areas. This enables real-time species recognition on basic hardware, aiding conservation efforts through rapid analysis. The model advances deep learning in ecology by balancing efficiency with performance.

Faces in Event Streams (FES): An Annotated Face Dataset for Event Cameras.

The use of event-based cameras in computer vision is a growing research direction. However, despite the existing research on face detection using the event camera, a substantial gap persists in the availability of a large dataset featuring annotations for faces and facial landmarks on event streams, thus hampering the development of applications in this direction. In this work, we address this issue by publishing the first large and varied dataset (Faces in Event Streams) with a duration of 689 min for face and facial landmark detection in direct event-based camera outputs. In addition, this article presents 12 models trained on our dataset to predict bounding box and facial landmark coordinates with an mAP50 score of more than 90%. We also performed a demonstration of real-time detection with an event-based camera using our models.

Using an improved lightweight YOLOv8 model for real-time detection of multi-stage apple fruit in complex orchard environments

For the purpose of monitoring apple fruits effectively throughout the entire growth period in smart orchards. A lightweight model named YOLOv8n-ShuffleNetv2-Ghost-SE was proposed. The ShuffleNetv2 basic modules and down-sampling modules were alternately connected, replacing the Backbone of YOLOv8n model. The Ghost modules replaced the Conv modules and the C2fGhost modules replaced the C2f modules in the Neck part of the YOLOv8n. ShuffleNetv2 reduced the memory access cost through channel splitting operations. The Ghost module combined linear and non-linear convolutions to reduce the network computation cost. The Wise-IoU (WIoU) replaced the CIoU for calculating the bounding box regression loss, which dynamically adjusted the anchor box quality threshold and gradient gain allocation strategy, optimizing the size and position of predicted bounding boxes. The Squeeze-and-Excitation (SE) was embedded in the Backbone and Neck part of YOLOv8n to enhance the representation ability of feature maps. The algorithm ensured high precision while having small model size and fast detection speed, which facilitated model migration and deployment. Using 9652 images validated the effectiveness of the model. The YOLOv8n-ShuffleNetv2-Ghost-SE model achieved Precision of 94.1%, Recall of 82.6%, mean Average Precision of 91.4%, model size of 2.6 MB, parameters of 1.18 M, FLOPs of 3.9 G, and detection speed of 39.37 fps. The detection speeds on the Jetson Xavier NX development board were 3.17 fps. Comparisons with advanced models including Faster R-CNN, SSD, YOLOv5s, YOLOv7‑tiny, YOLOv8s, YOLOv8n, MobileNetv3_small-Faster, MobileNetv3_small-Ghost, ShuflleNetv2-Faster, ShuflleNetv2-Ghost, ShuflleNetv2-Ghost-CBAM, ShuflleNetv2-Ghost-ECA, and ShuflleNetv2-Ghost-CA demonstrated that the method achieved smaller model and faster detection speed. The research can provide reference for the development of smart devices in apple orchards.

Evaluation of two deep learning-based approaches for detecting weeds growing in cabbage.

Machine vision-based precision weed management is a promising solution to substantially reduce herbicide input and weed control cost. The objective of this research was to compare two different deep learning-based approaches for detecting weeds in cabbage: (1) detecting weeds directly, and (2) detecting crops by generating the bounding boxes covering the crops and any green pixels outside the bounding boxes were deemed as weeds. The precision, recall, F1-score, mAP0.5, mAP0.5:0.95 of You Only Look Once (YOLO) v5 for detecting cabbage were 0.986, 0.979, 0.982, 0.995, and 0.851, respectively, while these metrics were 0.973, 0.985, 0.979, 0.993, and 0.906 for YOLOv8, respectively. However, none of these metrics exceeded 0.891 when detecting weeds. The reduced performances for directly detecting weeds could be attributed to the diverse weed species at varying densities and growth stages with different plant morphologies. A segmentation procedure demonstrated its effectiveness for extracting weeds outside the bounding boxes covering the crops, and thereby realizing effective indirect weed detection. The indirect weed detection approach demands less manpower as the need for constructing a large training dataset containing a variety of weed species is unnecessary. However, in a certain case, weeds are likely to remain undetected due to their growth in close proximity with crops and being situated within the predicted bounding boxes that encompass the crops. The models generated in this research can be used in conjunction with the machine vision subsystem of a smart sprayer or mechanical weeder. © 2024 Society of Chemical Industry.

RCDAM-Net: A Foreign Object Detection Algorithm for Transmission Tower Lines Based on RevCol Network

As an important part of the power system, it is necessary to ensure the safe and stable operation of transmission lines. Due to long-term exposure to the outdoors, the lines face many insecurity factors, and foreign object intrusion is one of them. Traditional foreign object (bird’s nest, kite, balloon, trash bag) detection algorithms suffer from low efficiency, poor accuracy, and small coverage, etc. To address the above problems, this paper introduces the RCDAM-Net. In order to prevent feature loss or useful feature compression, the RevCol (Reversible Column Networks) is used as the backbone network to ensure that the total information remains unchanged during feature decoupling. DySnakeConv (Dynamic Snake Convolution) is adopted and embedded into the C2f structure, which is named C2D and integrates low-level features and high-level features. Compared to the original BottleNeck structure of C2f, the DySnakeConv enhances the feature extraction ability for elongated and weak targets. In addition, MPDIoU (Maximum Performance Diagonal Intersection over Union) is used to improve the regression performance of model bounding boxes, solving the problem of predicted bounding boxes having the same aspect ratio as true bounding boxes, but with different values. Further, we adopt Decoupled Head for detection and add additional auxiliary training heads to improve the detection accuracy of the model. The experimental results show that the model achieves mAP50, Precision, and Recall of 97.98%, 98.15%, and 95.16% on the transmission tower line foreign object dataset, which is better to existing multi-target detection algorithms.

Protocol for metadata and image collection at diabetic foot ulcer clinics: enabling research in wound analytics and deep learning

BackgroundThe escalating impact of diabetes and its complications, including diabetic foot ulcers (DFUs), presents global challenges in quality of life, economics, and resources, affecting around half a billion people. DFU healing is hindered by hyperglycemia-related issues and diverse diabetes-related physiological changes, necessitating ongoing personalized care. Artificial intelligence and clinical research strive to address these challenges by facilitating early detection and efficient treatments despite resource constraints. This study establishes a standardized framework for DFU data collection, introducing a dedicated case report form, a comprehensive dataset named Zivot with patient population clinical feature breakdowns and a baseline for DFU detection using this dataset and a UNet architecture.ResultsFollowing this protocol, we created the Zivot dataset consisting of 269 patients with active DFUs, and about 3700 RGB images and corresponding thermal and depth maps for the DFUs. The effectiveness of collecting a consistent and clean dataset was demonstrated using a bounding box prediction deep learning network that was constructed with EfficientNet as the feature extractor and UNet architecture. The network was trained on the Zivot dataset, and the evaluation metrics showed promising values of 0.79 and 0.86 for F1-score and mAP segmentation metrics.ConclusionsThis work and the Zivot database offer a foundation for further exploration of holistic and multimodal approaches to DFU research.

Identifying the Growth Status of Hydroponic Lettuce Based on YOLO-EfficientNet

Hydroponic lettuce was prone to pest and disease problems after transplantation. Manual identification of the current growth status of each hydroponic lettuce not only consumed time and was prone to errors but also failed to meet the requirements of high-quality and efficient lettuce cultivation. In response to this issue, this paper proposed a method called YOLO-EfficientNet for identifying the growth status of hydroponic lettuce. Firstly, the video data of hydroponic lettuce were processed to obtain individual frame images. And 2240 images were selected from these frames as the image dataset A. Secondly, the YOLO-v8n object detection model was trained using image dataset A to detect the position of each hydroponic lettuce in the video data. After selecting the targets based on the predicted bounding boxes, 12,000 individual lettuce images were obtained by cropping, which served as image dataset B. Finally, the EfficientNet-v2s object classification model was trained using image dataset B to identify three growth statuses (Healthy, Diseases, and Pests) of hydroponic lettuce. The results showed that, after training image dataset A using the YOLO-v8n model, the accuracy and recall were consistently around 99%. After training image dataset B using the EfficientNet-v2s model, it achieved excellent scores of 95.78 for Val-acc, 94.68 for Test-acc, 96.02 for Recall, 96.32 for Precision, and 96.18 for F1-score. Thus, the method proposed in this paper had potential in the agricultural application of identifying and classifying the growth status in hydroponic lettuce.

Predicting diabetic macular edema in retina fundus images based on optimized deep residual network techniques on medical internet of things

With the medical internet of things, many automated diagnostic models related to eye diseases are easier. The doctors could quickly contrast and compare retina fundus images. The retina image contains a lot of information in the image. The task of detecting diabetic macular edema from retinal images in the healthcare system is difficult because the details in these images are very small. This paper proposed the new model based on the medical internet of things for predicting diabetic macular edema in retina fundus images. The method called DMER (Diabetic Macular Edema in Retina fundus images) to detect diabetic macular edema in retina fundus images based on improving deep residual network being combined with feature pyramid network in the context of the medical internet of things. The DMER method includes the following stages: (i) ResNet101 improved combining with feature pyramid network is used to extract features of the image and obtain the map of these features; (ii) a region proposal network to look for potential anomalies; and (iii) the predicted bounding boxes against the true bounding box by the regression method to certify the capability of macular edema. The MESSIDOR and DIARETDB1 datasets are used for testing with evaluation criteria such as sensitivity, specificity, and accuracy. The accuracy of the DMER method is about 98.08% with MESSIDOR dataset and 98.92% with DIARETDB1 dataset. The results of the method DMER are better than those of the other methods up to the present time with the above datasets.

Automatic Detection of Feral Pigeons in Urban Environments Using Deep Learning.

The overpopulation of feral pigeons in Hong Kong has significantly disrupted the urban ecosystem, highlighting the urgent need for effective strategies to control their population. In general, control measures should be implemented and re-evaluated periodically following accurate estimations of the feral pigeon population in the concerned regions, which, however, is very difficult in urban environments due to the concealment and mobility of pigeons within complex building structures. With the advances in deep learning, computer vision can be a promising tool for pigeon monitoring and population estimation but has not been well investigated so far. Therefore, we propose an improved deep learning model (Swin-Mask R-CNN with SAHI) for feral pigeon detection. Our model consists of three parts. Firstly, the Swin Transformer network (STN) extracts deep feature information. Secondly, the Feature Pyramid Network (FPN) fuses multi-scale features to learn at different scales. Lastly, the model's three head branches are responsible for classification, best bounding box prediction, and segmentation. During the prediction phase, we utilize a Slicing-Aided Hyper Inference (SAHI) tool to focus on the feature information of small feral pigeon targets. Experiments were conducted on a feral pigeon dataset to evaluate model performance. The results reveal that our model achieves excellent recognition performance for feral pigeons.

Task-Wise Sampling Convolutions for Arbitrary-Oriented Object Detection in Aerial Images.

Arbitrary-oriented object detection (AOOD) has been widely applied to locate and classify objects with diverse orientations in remote sensing images. However, the inconsistent features for the localization and classification tasks in AOOD models may lead to ambiguity and low-quality object predictions, which constrains the detection performance. In this article, an AOOD method called task-wise sampling convolutions (TS-Conv) is proposed. TS-Conv adaptively samples task-wise features from respective sensitive regions and maps these features together in alignment to guide a dynamic label assignment for better predictions. Specifically, sampling positions of the localization convolution in TS-Conv are supervised by the oriented bounding box (OBB) prediction associated with spatial coordinates, while sampling positions and convolutional kernel of the classification convolution are designed to be adaptively adjusted according to different orientations for improving the orientation robustness of features. Furthermore, a dynamic task-consistent-aware label assignment (DTLA) strategy is developed to select optimal candidate positions and assign labels dynamically according to ranked task-aware scores obtained from TS-Conv. Extensive experiments on several public datasets covering multiple scenes, multimodal images, and multiple categories of objects demonstrate the effectiveness, scalability, and superior performance of the proposed TS-Conv.

A Method for Training Object Scale Estimation System using Feature Extraction Enhancement with Depth Estimation

In recent years, machine learning-based object scale estimation has been growing in popularity, as the significance of the technology lies in its potential for use in many industry fields. Although several methods have been proposed, the possible applications of this technique are limited due to its insufficient accuracy. Hence, a human-level accurate system is needed for the technology to be applied in the real-world domain. This research paper proposes a novel object scale estimation system that incorporates the feature extractor, disentangled feature maps, depth estimator, object localizer, and ground truth depth map. The input of the proposed system is an image, which is inputted into the feature extractor to create disentangled feature maps. These feature maps are then extracted by the depth estimator to generate a depth map, and by the object localizer to create a predicted bounding box around the object. The trained feature extractor can extract disentangled size-related features from the inputted image by jointly training the depth estimator and object localizer. The use of disentangled features boosts the performance of the proposed system. In addition, we propose an actual scale converter module to calculate the actual size of the inputted object. Throughout the experiments, the proposed method has proven that it is superior compared to other state-of-the-art methods. The proposed method achieves an IoU (Intersection over Union) value of 0.8113 on the COCO dataset.

Elevating healthcare through artificial intelligence: analyzing the abdominal emergencies data set (TR_ABDOMEN_RAD_EMERGENCY) at TEKNOFEST-2022.

The artificial intelligence competition in healthcare at TEKNOFEST-2022 provided a platform to address the complex multi-class classification challenge of abdominal emergencies using computer vision techniques. This manuscript aimed to comprehensively present the methodologies for data preparation, annotation procedures, and rigorous evaluation metrics. Moreover, it was conducted to introduce a meticulously curated abdominal emergencies data set to the researchers. The data set underwent a comprehensive central screening procedure employing diverse algorithms extracted from the e-Nabız (Pulse) and National Teleradiology System of the Republic of Türkiye, Ministry of Health. Full anonymization of the data set was conducted. Subsequently, the data set was annotated by a group of ten experienced radiologists. The evaluation process was executed by calculating F1 scores, which were derived from the intersection over union values between the predicted bounding boxes and the corresponding ground truth (GT) bounding boxes. The establishment of baseline performance metrics involved computing the average of the highest five F1 scores. Observations indicated a progressive decline in F1 scores as the threshold value increased. Furthermore, it could be deduced that class 6 (abdominal aortic aneurysm/dissection) was relatively straightforward to detect compared to other classes, with class 5 (acute diverticulitis) presenting the most formidable challenge. It is noteworthy, however, that if all achieved outcomes for all classes were considered with a threshold of 0.5, the data set's complexity and associated challenges became pronounced. This data set's significance lies in its pioneering provision of labels and GT-boxes for six classes, fostering opportunities for researchers. The prompt identification and timely intervention in cases of emergent medical conditions hold paramount significance. The handling of patients' care can be augmented, while the potential for errors is minimized, particularly amidst high caseload scenarios, through the application of AI. • The data set used in artificial intelligence competition in healthcare (TEKNOFEST-2022) provides a 6-class data set of abdominal CT images consisting of a great variety of abdominal emergencies. • This data set is compiled from the National Teleradiology System data repository of emergency radiology departments of 459 hospitals. • Radiological data on abdominal emergencies is scarce in literature and this annotated competition data set can be a valuable resource for further studies and new AI models.

BFE-Net: Object Detection with Bidirectional Feature Enhancement

In realistic scenarios, existing object detection models still face challenges in resisting interference and detecting small objects due to complex environmental factors such as light and noise. For this reason, a novel scheme termed BFE-Net based on bidirectional feature enhancement is proposed. Firstly, a new multi-scale feature extraction module is constructed, which uses a self-attention mechanism to simulate human visual perception. It is used to capture global information and long-range dependencies between pixels, thereby optimizing the extraction of multi-scale features from input images. Secondly, a feature enhancement and denoising module is designed, based on bidirectional information flow. In the top-down, the impact of noise on the feature map is weakened to further enhance the feature extraction. In the bottom-up, multi-scale features are fused to improve the accuracy of small object feature extraction. Lastly, a generalized intersection over union regression loss function is employed to optimize the movement direction of predicted bounding boxes, improving the efficiency and accuracy of object localization. Experimental results using the public dataset PASCAL VOC2007test show that our scheme achieves a mean average precision (mAP) of 85% for object detection, which is 2.3% to 8.6% higher than classical methods such as RetinaNet and YOLOv5. Particularly, the anti-interference capability and the performance in detecting small objects show a significant enhancement.

An Automated Method for Generating Prefabs of AR Map Point Symbols Based on Object Detection Model

Augmented reality (AR) technology enables paper maps to dynamically express three-dimensional geographic information, realizing the fusion of virtual and real information. However, in the current mainstream AR development software, the virtual information usually consists of prefabricated components (prefabs), and the content creation for AR maps heavily relies on manual prefabrication. It leads to repetitive and error-prone prefabrication work, which restricts the design of the dynamic, interactive functions of AR maps. To solve this problem, this paper explored the possibility of automatically generating AR map prefabs using object detection models to establish a data conversion interface from paper maps to AR maps. First, we compared and analyzed various object detection models and selected YOLOv8x to recognize map point symbols. Then, we proposed a method to automatically generate AR map prefabs based on the predicted bounding boxes of the object detection model, which could generate prefabs with corresponding categories and positional information. Finally, we developed an AR map prototype system based on Android mobile devices. We designed an interaction method for information queries in the system to verify the effectiveness of the method proposed in this paper. The validation results indicate that our method can be practically applied to the AR map prefabrication process and can quickly generate AR map prefabs with high information accuracy. It alleviated the repetitive workload established through the manual prefabrication method and had specific feasibility and practicality. Moreover, it could provide solid data support for developing dynamic interactive functions of AR maps.

Spatio-temporal hierarchical feature transformer for UAV object tracking

Most existing Siamese network based trackers calculate the similarity between template and search region through a simple correlation operation, and these trackers classify and regress the target object based on the similarity maps obtained by the correlation operation. However, the correlation operation itself is a local linear matching process, if these Siamese network based trackers decide only on feature maps obtained by the correlation operation, using only a single similarity feature map will reduce position accuracy in challenging environments, using multiple similarity feature maps separately will introduce complicated computations. Thus, we propose an efficient spatio-temporal hierarchical feature transformer (STHFT) for UAV tracking. The spatio-temporal hierarchical similarity maps generated by the multi-level convolutional layers are sent to the transformer to achieve the interactive fusion of shallow spatio-temporal and deep semantic information. Therefore, our STHFT not only enhances the learning ability for global contextual information, but also effectively learns the interdependencies among multi-level features. Moreover, STHFT treats object tracking as a straightforward bounding box prediction task, only using a simple fully convolutional neural network to directly estimate the corners of the target and localize it. We have conducted extensive experiments on the UAV123 and UAV20L benchmarks, STHFT outperforms other traditional trackers. Specifically, compared to the conventional tracker DaSiamRPN the success and accuracy scores of STHFT improved by 7.3% and 6.5% on the UAV123 dataset and 10.9% and 10.4% on UAV20L, which show that STHFT is robust to tracking challenges.

Simultaneous object detection and segmentation for patient-specific markerless lung tumor tracking in simulated radiographs with deep learning.

Real-time tumor tracking is one motion management method to address motion-induced uncertainty. To date, fiducial markers are often required to reliably track lung tumors with X-ray imaging, which carries risks of complications and leads to prolonged treatment time. A markerless tracking approach is thus desirable. Deep learning-based approaches have shown promise for markerless tracking, but systematic evaluation and procedures to investigate applicability in individual cases are missing. Moreover, few efforts have been made to provide bounding box prediction and mask segmentation simultaneously, which could allow either rigid or deformable multi-leaf collimatortracking. The purpose of this study was to implement a deep learning-based markerless lung tumor tracking model exploiting patient-specific training which outputs both a bounding box and a mask segmentation simultaneously. We also aimed to compare the two kinds of predictions and to implement a specific procedure to understand the feasibility of markerless tracking on individualcases. We first trained a Retina U-Net baseline model on digitally reconstructed radiographs (DRRs) generated from a public dataset containing 875 CT scans and corresponding lung nodule annotations. Afterwards, we used an independent cohort of 97 lung patients to develop a patient-specific refinement procedure. In order to determine the optimal hyperparameters for automatic patient-specific training, we selected 13 patients for validation where the baseline model predicted a bounding box on planning CT (PCT)-DRR with intersection over union (IoU) with the ground-truth higher than 0.7. The final test set contained the remaining 84 patients with varying PCT-DRR IoU. For each testing patient, the baseline model was refined on the PCT-DRR to generate a patient-specific model, which was then tested on a separate 10-phase 4DCT-DRR to mimic the intrafraction motion during treatment. A template matching algorithm served as benchmark model. The testing results were evaluated by four metrics: the center of mass (COM) error and the Dice similarity coefficient (DSC) for segmentation masks, and the center of box (COB) error and the DSC for bounding box detections. Performance was compared to the benchmark model including statistical testing forsignificance. A PCT-DRR IoU value of 0.2 was shown to be the threshold dividing inconsistent (68%) and consistent (100%) success (defined as mean bounding box DSC > 0.6) of PS models on 4DCT-DRRs. Thirty-seven out of the eighty-four testing cases had a PCT-DRR IoU above 0.2. For these 37 cases, the mean COM error was 2.6 mm, the mean segmentation DSC was 0.78, the mean COB error was 2.7 mm, and the mean box DSC was 0.83. Including the validation cases, the model was applicable to 50 out of 97 patients when using the PCT-DRR IoU threshold of 0.2. The inference time per frame was 170 ms. The model outperformed the benchmark model on all metrics, and the comparison was significant (p < 0.001) over the 37 PCT-DRR IoU > 0.2 cases, but not over the undifferentiated 84 testingcases. The implemented patient-specific refinement approach based on a pre-trained baseline model was shown to be applicable to markerless tumor tracking in simulated radiographs for lungcases.

A Study on Tomato Disease and Pest Detection Method

In recent years, with the rapid development of artificial intelligence technology, computer vision-based pest detection technology has been widely used in agricultural production. Tomato diseases and pests are serious problems affecting tomato yield and quality, so it is important to detect them quickly and accurately. In this paper, we propose a tomato disease and pest detection model based on an improved YOLOv5n to overcome the problems of low accuracy and large model size in traditional pest detection methods. Firstly, we use the Efficient Vision Transformer as the feature extraction backbone network to reduce model parameters and computational complexity while improving detection accuracy, thus solving the problems of poor real-time performance and model deployment. Second, we replace the original nearest neighbor interpolation upsampling module with the lightweight general-purpose upsampling operator Content-Aware ReAssembly of FEatures to reduce feature information loss during upsampling. Finally, we use Wise-IoU instead of the original CIoU as the regression loss function of the target bounding box to improve the regression prediction accuracy of the predicted bounding box while accelerating the convergence speed of the regression loss function. We perform statistical analysis on the experimental results of tomato diseases and pests under data augmentation conditions. The results show that the improved algorithm improves mAP50 and mAP50:95 by 2.3% and 1.7%, respectively, while reducing the number of model parameters by 0.4 M and the computational complexity by 0.9 GFLOPs. The improved model has a parameter count of only 1.6 M and a computational complexity of only 3.3 GFLOPs, demonstrating a certain advantage over other mainstream object detection algorithms in terms of detection accuracy, model parameter count, and computational complexity. The experimental results show that this method is suitable for the early detection of tomato diseases and pests.