Mean Average Precision Score Research Articles

VisioSignNet: A Dual-Interactive Neural Network for enhanced traffic sign detection

Effective traffic sign detection is crucial for the safety and operational efficiency of autonomous vehicle navigation systems, particularly in dynamically changing environments. Addressing the primary challenges of long-range pixel dependencies and enhancing the detectability of small objects in complex scenes, we present VisioSignNet: A Dual-Interactive Neural Network designed for enhanced traffic sign detection. This architecture incorporates Local and Global Interactive Modules (LGIM) and Enhancing Channel and Space Interaction (ECSI) modules. The LGIM is engineered to balance local and global feature interactions, while the ECSI optimizes the interchange of information across channel and spatial dimensions. Their synergistic interaction not only enhances the perceptual field at early processing stages but also significantly improves the recognition of small-scale, critical traffic signs. Evaluated on the TT100K and GTSDB datasets, VisioSignNet achieved mean average precision (mAP) scores of 90.5% and 97%, respectively, with a model size of 26M parameters. Its enhanced variant, VisioSignNet_l, with 34M parameters, reached mAP scores of 93.2% and 97.8%. These outcomes substantiate VisioSignNet’s efficacy in tackling the complexities of traffic sign detection, confirming its potential as a robust solution in the field of autonomous driving technologies.

XRAInet: AI-based decision support for pneumothorax and pleural effusion management.

This study aimed to develop and assess the performance of an artificial intelligence(AI)-driven decision support system, XRAInet, in accurately identifying pediatric patients with pleural effusion or pneumothorax and determining whether tube thoracostomy intervention is warranted. In this diagnostic accuracy study, we retrospectively analyzed a data set containing 510 X-ray images from 170 pediatric patients admitted between 2005 and 2022. Patients were categorized into two groups: Tube (requiring tube thoracostomy) and Conservative (managed conservatively). XRAInet, a deep learning-based algorithm, was trained using this data set. We evaluated its performance using various metrics, including mean Average Precision (mAP), recall, precision, and F1 score. XRAInet, achieved a mAP score of 0.918. This result underscores its ability to accurately identify and localize regions necessitating tube thoracostomy for pediatric patients with pneumothorax and pleural effusion. In an independent testing data set, the model exhibited a sensitivity of 64.00% and specificity of 96.15%. In conclusion, XRAInet presents a promising solution for improving the detection and decision-making process for cases of pneumothorax and pleural effusion in pediatric patients using X-ray images. These findings contribute to the expanding field of AI-driven medical imaging, with potential applications for enhancing patient outcomes. Future research endeavors should explore hybrid models, enhance interpretability, address data quality issues, and align with regulatory requirements to ensure the safe and effective deployment of XRAInet in healthcare settings.

DeepLeish: a deep learning based support system for the detection of Leishmaniasis parasite from Giemsa-stained microscope images

BackgroundLeishmaniasis is a vector-born neglected parasitic disease belonging to the genus Leishmania. Out of the 30 Leishmania species, 21 species cause human infection that affect the skin and the internal organs. Around, 700,000 to 1,000,000 of the newly infected cases and 26,000 to 65,000 deaths are reported worldwide annually. The disease exhibits three clinical presentations, namely, the cutaneous, muco-cutaneous and visceral Leishmaniasis which affects the skin, mucosal membrane and the internal organs, respectively. The relapsing behavior of the disease limits its diagnosis and treatment efficiency. The common diagnostic approaches follow subjective, error-prone, repetitive processes. Despite, an ever pressing need for an accurate detection of Leishmaniasis, the research conducted so far is scarce. In this regard, the main aim of the current research is to develop an artificial intelligence based detection tool for the Leishmaniasis from the Geimsa-stained microscopic images using deep learning method.MethodsStained microscopic images were acquired locally and labeled by experts. The images were augmented using different methods to prevent overfitting and improve the generalizability of the system. Fine-tuned Faster RCNN, SSD, and YOLOV5 models were used for object detection. Mean average precision (MAP), precision, and Recall were calculated to evaluate and compare the performance of the models.ResultsThe fine-tuned YOLOV5 outperformed the other models such as Faster RCNN and SSD, with the MAP scores, of 73%, 54% and 57%, respectively.ConclusionThe currently developed YOLOV5 model can be tested in the clinics to assist the laboratorists in diagnosing Leishmaniasis from the microscopic images. Particularly, in low-resourced healthcare facilities, with fewer qualified medical professionals or hematologists, our AI support system can assist in reducing the diagnosing time, workload, and misdiagnosis. Furthermore, the dataset collected by us will be shared with other researchers who seek to improve upon the detection system of the parasite. The current model detects the parasites even in the presence of the monocyte cells, but sometimes, the accuracy decreases due to the differences in the sizes of the parasite cells alongside the blood cells. The incorporation of cascaded networks in future and the quantification of the parasite load, shall overcome the limitations of the currently developed system.

VTHSC-MIR: Vision Transformer Hashing with Supervised Contrastive learning based medical image retrieval

In past few years, deep learning based medical image analysis technologies have significantly improved computer-assisted tasks like detecting, diagnosing, and predicting medical outcomes. The monitoring and diagnosis of ailments such as cancer and COVID-19 rely primarily on retrieving medical images. As the medical databases size is increasing rapidly it causes difficulty in managing and querying them. The main challenge lies in achieving superior retrieval performance while handling the intricacies of medical imaging datasets. To tackle this, we propose a novel end-to-end trainable framework for medical image retrieval using a vision transformer hashing with Supervised Contrastive (VTHSC) learning. Leveraging the self-attention mechanism of transformers and supervised contrastive loss, our model surpasses existing hashing-based retrieval methods. Focusing on the most recent COVID-19 and breast cancer datasets, we aim to efficiently retrieve relevant medical images from large datasets. Using the ImageNet-pretrained ViT as its backbone network, the VTHSC model incorporates a hashing head combined with a supervised contrastive loss and employs joint loss optimization. The VTHSC model is subsequently fine- tuned for a retrieval task, employing four different hashing frameworks: Deep Supervised Hashing (DSH), GreedyHash, Improved Deep Hashing Network (IDHN), and Deep Polarized Network (DPN). Our model achieves outstanding Mean Average Precision (MAP) scores of 98.9 for the BreakHis dataset and 96.03 for the COVID dataset. Notably, the VTHSC model surpasses several competing hashing-based retrieval methods by a substantial gain in terms of performance across various metrics such as precision, recall, and Top-6 retrieved images retrieval performance on the two benchmark medical image datasets.

GDB‐YOLOv5s: Improved YOLO‐based model for ship detection in SAR images

AbstractIn recent years, deep learning methods were good solutions for object detection in synthetic aperture radar (SAR) images. However, the problems of complex scenarios, large object scale differences and imperfect fine‐grained classification in ship detection were intractable. In response, an improved model GDB‐YOLOv5s (Improved YOLOv5s model incorporating global attention mechanism (GAM), DCN‐v2 and BiFusion) is designed. This model introduces deformable convolution networks (DCN‐v2) into the Backbone to enhance the adaptability of the receptive field. It replaces the original Neck's PANet structure with a BiFusion structure to better fuse the extracted multiscale features. Additionally, it integrates GAM into the network to reduce information loss and improve global feature interaction. Experiments were conducted on single‐class dataset SSDD and multi‐class dataset SRSSD‐V1.0. The results show that the GDB‐YOLOv5s model improves mean average precision scores (mAP) significantly, outperforming the original YOLOv5s model and other traditional methods. GDB‐YOLOv5s also improves Precision‐score (P) and Recall‐score (R) for fine‐grained classification to some extent, thereby reducing false alarms and missed detections. It has been proved that the improved model is relatively effective.

DNN-PolSAR: Urban Image Segmentation and Classification using Polarimetric SAR based on DNNs

Synthetic Aperture Radar (SAR) image segmentation and classification is a popular technique for learning and detection of objects such as buildings, trees, monuments, crops water-bodies, hills, etc. SAR technique is being used for urban development and city-planning, building control of municipal objects, searching best locations, detection of changes in the existing systems, etc. using polarimetry based on Deep Neural Networks. In this paper, weproposed a technique for Urban Image Segmentation and Classification using Polarimetric SAR based on Deep NeuralNetworks (DNN-PolSAR). In our proposed DNN-PolSAR technique, we useMask-RCNN, LinkNet, FPN, and PSP- Net as model architectures, whereas ResNet50, ResNet101, ResNet152, and VGG-19 are used as backbone networks.We first apply polarimetric decomposition on airborne Uninhabited Aerial Vehicle Synthetic Aperture (UAVSAR) im- ages of urban areas and then the decomposed images are fed to DNNs for segmentation and classification. We then simulate DNN-PolSAR considering different hyper-parameters and compare the obtained scores of hyper-parametersagainst used model architectures and backbone networks. In comparison, it is found that DNN-PolSAR based on FPN model with ResNet152 performed the best for segmentation and classification. The mean Average Precision (mAP) score of the DNN-PolSAR based on FPN with a pixel accuracy of 90.9% is 0.823, which outperforms other Deep Learning models.

A Cross-Language Information Retrieval Method Based on Multi-Task Learning

This study introduces a novel Cross-Language Information Retrieval (CLIR) method employing multi-task learning and soft parameter sharing to enhance neural retrieval models' feature extraction across languages. The approach integrates an interaction-based neural retrieval model with a semantic-based text classification model, exchanging hidden vectors for richer feature representation. Experimental results across four language pairs—English-Chinese, English-Arabic, English-French, and English-German—demonstrate significant performance improvements. The proposed method achieved the highest Mean Average Precision (MAP) scores: 0.419 for EN-ZH, 0.403 for EN-AR, 0.427 for EN-FR, and 0.441 for EN-DE, surpassing other models like BM25, BPNRM, KNRM, KNRM-Trans, and KNRM-Embed. This research underscores the potential of multi-task learning for CLIR, showcasing improved retrieval performance through semantic information and knowledge transfer.

A Comparative Dataset of Annotated Broccoli Heads Recorded with Depth Cameras from a Moving Vehicle

An extensive, publicly available dataset is presented—the LAR Broccoli dataset—which contains 20,000 manually annotated images of broccoli heads captured from a moving tractor at an organic farm in the UK. The dataset contains images of the same row of broccoli heads recorded at 30 frames per second (fps) with three different cameras. Two off-the-shelf, relatively low-cost depth-sensing cameras were used, with the tractor moving at a speed of around 1 km/h, in addition to a webcam, with the tractor moving twice as fast. The utility of the dataset is demonstrated in four ways. First, three different state-of-the-art detector models were trained on the dataset, achieving an overall mean Average Precision (mAP) score of over 95% for the best-performing detector. The results validate the utility of the dataset for the standard task of in-field broccoli head recognition. Second, experiments with transfer learning were conducted, initialised with a smaller pre-trained broccoli detection model, and refined with the LAR Broccoli dataset. Third, we assessed the advantages of transfer learning not only using mAP but also according to time and space requirements for training models, which provides a proxy metric for energy efficiency, a practical consideration for real-world model training. Fourth, the cross-camera generalisation among the three camera systems was compared. The results highlight that testing and training detector models using different camera systems can lead to reduced performance, unless the training set also includes some images captured in the same manner as those in the test set.

Enabling near-real-time safety glove detection through edge computing and transfer learning: comparative analysis of edge and cloud computing-based methods

PurposeComputer vision and deep learning (DL) methods have been investigated for personal protective equipment (PPE) monitoring and detection for construction workers’ safety. However, it is still challenging to implement automated safety monitoring methods in near real time or in a time-efficient manner in real construction practices. Therefore, this study developed a novel solution to enhance the time efficiency to achieve near-real-time safety glove detection and meanwhile preserve data privacy.Design/methodology/approachThe developed method comprises two primary components: (1) transfer learning methods to detect safety gloves and (2) edge computing to improve time efficiency and data privacy. To compare the developed edge computing-based method with the currently widely used cloud computing-based methods, a comprehensive comparative analysis was conducted from both the implementation and theory perspectives, providing insights into the developed approach’s performance.FindingsThree DL models achieved mean average precision (mAP) scores ranging from 74.92% to 84.31% for safety glove detection. The other two methods by combining object detection and classification achieved mAP as 89.91% for hand detection and 100% for glove classification. From both implementation and theory perspectives, the edge computing-based method detected gloves faster than the cloud computing-based method. The edge computing-based method achieved a detection latency of 36%–68% shorter than the cloud computing-based method in the implementation perspective. The findings highlight edge computing’s potential for near-real-time detection with improved data privacy.Originality/valueThis study implemented and evaluated DL-based safety monitoring methods on different computing infrastructures to investigate their time efficiency. This study contributes to existing knowledge by demonstrating how edge computing can be used with DL models (without sacrificing their performance) to improve PPE-glove monitoring in a time-efficient manner as well as maintain data privacy.

YOLO-Based Models for Smoke and Wildfire Detection in Ground and Aerial Images

Wildland fires negatively impact forest biodiversity and human lives. They also spread very rapidly. Early detection of smoke and fires plays a crucial role in improving the efficiency of firefighting operations. Deep learning techniques are used to detect fires and smoke. However, the different shapes, sizes, and colors of smoke and fires make their detection a challenging task. In this paper, recent YOLO-based algorithms are adopted and implemented for detecting and localizing smoke and wildfires within ground and aerial images. Notably, the YOLOv7x model achieved the best performance with an mAP (mean Average Precision) score of 80.40% and fast detection speed, outperforming the baseline models in detecting both smoke and wildfires. YOLOv8s obtained a high mAP of 98.10% in identifying and localizing only wildfire smoke. These models demonstrated their significant potential in handling challenging scenarios, including detecting small fire and smoke areas; varying fire and smoke features such as shape, size, and colors; the complexity of background, which can include diverse terrain, weather conditions, and vegetation; and addressing visual similarities among smoke, fog, and clouds and the the visual resemblances among fire, lighting, and sun glare.

Incremental convolutional transformer for baggage threat detection

Detecting cluttered and overlapping contraband items from baggage scans is one of the most challenging tasks, even for human experts. Recently, considerable literature has grown up around the theme of deep learning-based X-ray screening for localizing contraband data. However, the existing threat detection systems are still vulnerable to high occlusion, clutter, and concealment. Furthermore, they require exhaustive training routines on large-scale and well-annotated data in order to produce accurate results. To overcome the above-mentioned limitations, this paper presents a novel convolutional transformer system that recognizes different overlapping instances of prohibited objects in complex baggage X-ray scans via a distillation-driven incremental instance segmentation scheme. Furthermore, unlike its competitors, the proposed framework allows an incremental integration of new item instances while avoiding costly training routines. In addition to this, the proposed framework also outperforms state-of-the-art approaches by achieving a mean average precision score of 0.7896, 0.5974, and 0.7569 on publicly available GDXray, SIXray, and OPIXray datasets for detecting concealed and cluttered baggage threats.

RETRACTED: Utilizing Generative Adversarial Networks for Acne Dataset Generation in Dermatology

Background: In recent years, computer-aided diagnosis for skin conditions has made significant strides, primarily driven by artificial intelligence (AI) solutions. However, despite this progress, the efficiency of AI-enabled systems remains hindered by the scarcity of high-quality and large-scale datasets, primarily due to privacy concerns. Methods: This research circumvents privacy issues associated with real-world acne datasets by creating a synthetic dataset of human faces with varying acne severity levels (mild, moderate, and severe) using Generative Adversarial Networks (GANs). Further, three object detection models—YOLOv5, YOLOv8, and Detectron2—are used to evaluate the efficacy of the augmented dataset for detecting acne. Results: Integrating StyleGAN with these models, the results demonstrate the mean average precision (mAP) scores: YOLOv5: 73.5%, YOLOv8: 73.6%, and Detectron2: 37.7%. These scores surpass the mAP achieved without GANs. Conclusions: This study underscores the effectiveness of GANs in generating synthetic facial acne images and emphasizes the importance of utilizing GANs and convolutional neural network (CNN) models for accurate acne detection.

Fine-Grained Analysis of Coral Instance Segmentation using YOLOv8 Models

Within the geographical boundaries of Indonesia, coral reefs flourish as intricate ecosystems bustling with a variety of marine creatures that play a crucial role, in preserving biodiversity. However this delicate harmony faces threats from climate change and human activities, leading to the risk of species loss. Despite growing awareness surrounding these challenges effectively and swiftly monitoring conditions remains a task. Existing methods for assessing corals often fall short due to requiring extensive specialist knowledge, lacking large-scale coverage, and being costly to implement. To tackle these obstacles this research suggests an approach for automated reef monitoring using instance segmentation with a YOLOv8 model. Leveraging YOLOv8 segmentation capabilities enables efficient analysis of corals. A systematic process is employed involving data collection, preparation (including techniques like Histogram Equalization), training the model on a reef dataset, model evaluation and enhancing the segmentation mask. The outcomes reveal the YOLOv8m Pp model with 96.7% precision 95.9% recall rate and a mean Average Precision (mAP50) score of 98.2%. This study demonstrates the potential of YOLOv8 to accurately segment instances for monitoring reefs in Indonesia, hence facilitating improved conservation strategies.

Deep learning-based computer-aided diagnosis system for the automatic detection and classification of lateral cervical lymph nodes on original ultrasound images of papillary thyroid carcinoma: a prospective diagnostic study.

This study aims to develop a deep learning-based computer-aided diagnosis (CAD) system for the automatic detection and classification of lateral cervical lymph nodes (LNs) on original ultrasound images of papillary thyroid carcinoma (PTC) patients. A retrospective data set of 1801 cervical LN ultrasound images from 1675 patients with PTC and a prospective test set including 185 images from 160 patients were collected. Four different deep leaning models were trained and validated in the retrospective data set. The best model was selected for CAD system development and compared with three sonographers in the retrospective and prospective test sets. The Deformable Detection Transformer (DETR) model showed the highest diagnostic efficacy, with a mean average precision score of 86.3% in the retrospective test set, and was therefore used in constructing the CAD system. The detection performance of the CAD system was superior to the junior sonographer and intermediate sonographer with accuracies of 86.3% and 92.4% in the retrospective and prospective test sets, respectively. The classification performance of the CAD system was better than all sonographers with the areas under the curve (AUCs) of 94.4% and 95.2% in the retrospective and prospective test sets, respectively. This study developed a Deformable DETR model-based CAD system for automatically detecting and classifying lateral cervical LNs on original ultrasound images, which showed excellent diagnostic efficacy and clinical utility. It can be an important tool for assisting sonographers in the diagnosis process.

Diabetic Foot Ulcer Detection: Combining Deep Learning Models for Improved Localization

Diabetes mellitus (DM) can cause chronic foot issues and severe infections, including Diabetic Foot Ulcers (DFUs) that heal slowly due to insufficient blood flow. A recurrence of these ulcers can lead to 84% of lower limb amputations and even cause death. High-risk diabetes patients require expensive medications, regular check-ups, and proper personal hygiene to prevent DFUs, which affect 15–25% of diabetics. Accurate diagnosis, appropriate care, and prompt response can prevent amputations and fatalities through early and reliable DFU detection from image analysis. We propose a comprehensive deep learning-based system for detecting DFUs from patients’ feet images by reliably localizing ulcer points. Our method utilizes innovative model ensemble techniques—non-maximum suppression (NMS), Soft-NMS, and weighted bounding box fusion (WBF)—to combine predictions from state-of-the-art object detection models. The performances of diverse cutting-edge model architectures used in this study complement each other, leading to more generalized and improved results when combined in an ensemble. Our WBF-based approach combining YOLOv8m and FRCNN-ResNet101 achieves a mean average precision (mAP) score of 86.4% at the IoU threshold of 0.5 on the DFUC2020 dataset, significantly outperforming the former benchmark by 12.4%. We also perform external validation on the IEEE DataPort Diabetic Foot dataset which has demonstrated robust and reliable model performance on the qualitative analysis. In conclusion, our study effectively developed an innovative diabetic foot ulcer (DFU) detection system using an ensemble model of deep neural networks (DNNs). This AI-driven tool serves as an initial screening aid for medical professionals, augmenting the diagnostic process by enhancing sensitivity to potential DFU cases. While recognizing the presence of false positives, our research contributes to improving patient care through the integration of human medical expertise with AI-based solutions in DFU management.

An optimal model for detection of lung cancer using convolutional neural network

In terms of frequency and mortality, lung cancer ranks second among all cancers worldwide for both men and women. It is suggested that pattern classification and machine learning be applied to the identification and categorization of lung cancer. Convolution neural network (CNN) techniques divide the input data into groups according to the distinctive characteristics of the input. Using a standard approach to analyze a large number of computed tomography images, early detection of lung cancer can save lives. The suggested effort is centered on identifying the precise type of cancer and making predictions about whether it is benign or aggressive. The deployment of proposed model is an attempt to improve the accuracy of the system. The proposed work showed an overall accuracy of 98.4% during the detection of lung cancer and 98.8% accuracy towards the prediction of specific type in the lung cancer. Mean average precision score of 97.17% and 98.75% test and validation respectively. 0.96, 0.93, and 0.95 for malignant test data.

Deep learning assisted fabrication of metallic components using the robotic wire arc additive manufacturing

PurposeBecause of the advantages of high deposition efficiency and low manufacturing cost compared with other additive technologies, robotic wire arc additive manufacturing (WAAM) technology has been widely applied for fabricating medium- to large-scale metallic components. The additive manufacturing (AM) method is a relatively complex process, which involves the workpiece modeling, conversion of the model file, slicing, path planning and so on. Then the structure is formed by the accumulated weld bead. However, the poor forming accuracy of WAAM usually leads to severe dimensional deviation between the as-built and the predesigned structures. This paper aims to propose a visual sensing technology and deep learning–assisted WAAM method for fabricating metallic structure, to simplify the complex WAAM process and improve the forming accuracy.Design/methodology/approachInstead of slicing of the workpiece modeling and generating all the welding torch paths in advance of the fabricating process, this method is carried out by adding the feature point regression branch into the Yolov5 algorithm, to detect the feature point from the images of the as-built structure. The coordinates of the feature points of each deposition layer can be calculated automatically. Then the welding torch trajectory for the next deposition layer is generated based on the position of feature point.FindingsThe mean average precision score of modified YOLOv5 detector is 99.5%. Two types of overhanging structures have been fabricated by the proposed method. The center contour error between the actual and theoretical is 0.56 and 0.27 mm in width direction, and 0.43 and 0.23 mm in height direction, respectively.Originality/valueThe fabrication of circular overhanging structures without using the complicate slicing strategy, turning table or other extra support verified the possibility of the robotic WAAM system with deep learning technology.

Automatic Identification of Pangolin Behavior Using Deep Learning Based on Temporal Relative Attention Mechanism.

With declining populations in the wild, captive rescue and breeding have become one of the most important ways to protect pangolins from extinction. At present, the success rate of artificial breeding is low, due to the insufficient understanding of the breeding behavior characteristics of pangolins. The automatic recognition method based on machine vision not only monitors for 24 h but also reduces the stress response of pangolins. This paper aimed to establish a temporal relation and attention mechanism network (Pangolin breeding attention and transfer network, PBATn) to monitor and recognize pangolin behaviors, including breeding and daily behavior. There were 11,476 videos including breeding behavior and daily behavior that were divided into training, validation, and test sets. For the training set and validation set, the PBATn network model had an accuracy of 98.95% and 96.11%, and a loss function value of 0.1531 and 0.1852. The model is suitable for a 2.40 m × 2.20 m (length × width) pangolin cage area, with a nest box measuring 40 cm × 30 cm × 30 cm (length × width × height) positioned either on the left or right side inside the cage. A spherical night-vision monitoring camera was installed on the cage wall at a height of 2.50 m above the ground. For the test set, the mean Average Precision (mAP), average accuracy, average recall, average specificity, and average F1 score were found to be higher than SlowFast, X3D, TANet, TSN, etc., with values of 97.50%, 99.17%, 97.55%, 99.53%, and 97.48%, respectively. The recognition accuracies of PBATn were 94.00% and 98.50% for the chasing and mounting breeding behaviors, respectively. The results showed that PBATn outperformed the baseline methods in all aspects. This study shows that the deep learning system can accurately observe pangolin breeding behavior and it will be useful for analyzing the behavior of these animals.

Tea yield estimation using UAV images and deep learning

Traditional estimation of tea yield significantly depends on fields observation, which is facing operational and management challenges due to increasing farm sizes and rising labor costs. Development of artificial intelligence such as deep learning (DL) and unmanned aerial vehicles (UAVs) provide opportunities where intelligent system of tea fields will be constructed. This study carried out the application of tea yield estimation model to practical field scenarios. A large dataset of UAV tea buds’ images was built, which contained 5899 images and 29,958 labelled tea buds. Based on the large dataset, YOLOv5 model was used to train, and CSPDarknet53 performed best over Swin Transformer and ConvNeXt as backbone. The mean average precision (mAP), precision, recall and F1 score of CSPDarknet53 were respectively 85.63%, 84.91%,72.19% and 78.00%. Meanwhile, the large database weakened the differences among models of different sizes. Squeeze-and-excitation (SE) block was inserted, and the mAP of all models with CSPDarknet53 and SE was over 85.00%, indicating models could be applied to tea fields. To reduce the number of repeated tea bud detection boxes in the prediction process, a second non-maximum suppression (NMS) filter was used with threshold value of 0.20. The yield of fresh tea leaves before and after picking was estimated according to the number of tea buds detected in different seasons and different types of tea plants. The yield of fresh tea leaves in spring before picking was estimated to 1223.22 kg/ha. This study is a complete report from model training to tea field application, and will advance the industrialization development of agriculture.

Enhancing identification for person search with multi-scale multi-grained representation learning

Person Search aims to simultaneously address Person Detection and Person Re-ID. There are various challenges in person search such as significant scale variations, occlusions, and partial instances. In this paper, we propose a Multi-Scale Multi-Grained (MSMG) sequential network for end-to-end person search, intended to alleviate these issues. To generate re-id representations robust to scale changes, MSMG leverages multi-scale RoI features and aggregates them with a proposed Multi-Scale feature Aggregation Encoder (MSAE). In this way, the aggregated multi-scale re-id features are enriched with more semantic information and detailed information, thereby being more discriminative for identification. Moreover, to produce re-id representations more robust to occlusions and partial instances, MSMG introduces a Multi-Grained feature Learning Decoder (MGLD) focused on multi-grained feature learning. MGLD adaptively decodes multi-grained re-id representations with more accurate semantic information through a regional deformable cross-attention module. Finally, the multi-scale multi-grained re-id representation substantially improves the identification accuracy under challenging cases. Through comprehensive experiments, we demonstrate that our method achieves state-of-the-art performance on two benchmark datasets. On the challenging PRW benchmark, MSMG obtains the best-reported results with a mean average precision (mAP) score of 61.3%.