Mean Average Precision Score Research Articles

Daedalus: Breaking Nonmaximum Suppression in Object Detection via Adversarial Examples.

This article demonstrates that nonmaximum suppression (NMS), which is commonly used in object detection (OD) tasks to filter redundant detection results, is no longer secure. Considering that NMS has been an integral part of OD systems, thwarting the functionality of NMS can result in unexpected or even lethal consequences for such systems. In this article, an adversarial example attack that triggers malfunctioning of NMS in OD models is proposed. The attack, namely, Daedalus, compresses the dimensions of detection boxes to evade NMS. As a result, the final detection output contains extremely dense false positives. This can be fatal for many OD applications, such as autonomous vehicles and surveillance systems. The attack can be generalized to different OD models, such that the attack cripples various OD applications. Furthermore, a way of crafting robust adversarial examples is developed by using an ensemble of popular detection models as the substitutes. Considering the pervasive nature of model reuse in real-world OD scenarios, Daedalus examples crafted based on an ensemble of substitutes can launch attacks without knowing the parameters of the victim models. The experimental results demonstrate that the attack effectively stops NMS from filtering redundant bounding boxes. As the evaluation results suggest, Daedalus increases the false positive rate in detection results to 99.9% and reduces the mean average precision scores to 0, while maintaining a low cost of distortion on the original inputs. It also demonstrates that the attack can be practically launched against real-world OD systems via printed posters.

Development and Optimization of Deep Learning Models for Weapon Detection in Surveillance Videos

Weapon detection in CCTV camera surveillance videos is a challenging task and its importance is increasing because of the availability and easy access of weapons in the market. This becomes a big problem when weapons go into the wrong hands and are often misused. Advances in computer vision and object detection are enabling us to detect weapons in live videos without human intervention and, in turn, intelligent decisions can be made to protect people from dangerous situations. In this article, we have developed and presented an improved real-time weapon detection system that shows a higher mean average precision (mAP) score and better inference time performance compared to the previously proposed approaches in the literature. Using a custom weapons dataset, we implemented a state-of-the-art Scaled-YOLOv4 model that resulted in a 92.1 mAP score and frames per second (FPS) of 85.7 on a high-performance GPU (RTX 2080TI). Furthermore, to achieve the benefits of lower latency, higher throughput, and improved privacy, we optimized our model for implementation on a popular edge-computing device (Jetson Nano GPU) with the TensorRT network optimizer. We have also performed a comparative analysis of the previous weapon detector with our presented model using different CPU and GPU machines that fulfill the purpose of this work, making the selection of model and computing device easier for the users for deployment in a real-time scenario. The analysis shows that our presented models result in improved mAP scores on high-performance GPUs (such as RTX 2080TI), as well as on low-cost edge computing GPUs (such as Jetson Nano) for weapon detection in live CCTV camera surveillance videos.

Improving query expansion using pseudo-relevant web knowledge for information retrieval

• Web knowledge-based query expansion technique uses the top N pseudo relevant web pages • Top N pseudo relevant web pages returned by search engines Google, Bing, and DuckDuckGo as data sources • Proposed weighting techniques have properly capture the term relationship while weighing the expansion terms • Experimental results show an improvement of 25.89% over the unexpanded queries of FIRE dataset In the field of information retrieval, query expansion (QE) has long been used as a technique to deal with the fundamental issue of word mismatch between a user’s query and the target information. In the context of the relationship between the query and expanded terms, existing weighting techniques often fail to appropriately capture the term-term relationship and term to the whole query relationship, resulting in low retrieval effectiveness. Our proposed QE approach addresses this by proposing three weighting models based on (1) tf-idf, (2) k-nearest neighbor (kNN) based cosine similarity , and (3) correlation score. Further, to extract the initial set of expanded terms, we use pseudo-relevant web knowledge consisting of the top N web pages returned by the three popular search engines namely, Google, Bing, and DuckDuckGo, in response to the original query. Among the three weighting models, tf-idf scores each of the individual terms obtained from the web content, kNN-based cosine similarity scores the expansion terms to obtain the term-term relationship, and correlation score weighs the selected expansion terms with respect to the whole query. The proposed model, called web knowledge based query expansion (WKQE), achieves an improvement of 25.89% on the Mean Average Precision (MAP) score and 30.83% on the Geometric Mean Average precision (GMAP) score over the unexpanded queries on the FIRE dataset. A comparative analysis of the WKQE techniques with other related approaches clearly shows significant improvement in the retrieval performance. We have also analyzed the effect of varying the number of pseudo-relevant documents and expansion terms on the retrieval effectiveness of the proposed model.

A System for a Real-Time Electronic Component Detection and Classification on a Conveyor Belt

The presented research addresses the real-time object detection problem with small and moving objects, specifically the surface-mount component on a conveyor. Detecting and counting small moving objects on the assembly line is a challenge. In order to meet the requirements of real-time applications, state-of-the-art electronic component detection and classification algorithms are implemented into powerful hardware systems. This work proposes a low-cost system with an embedded microcomputer to detect surface-mount components on a conveyor belt in real time. The system detects moving, packed, and unpacked surface-mount components. The system’s performance was experimentally investigated by implementing several object-detection algorithms. The system’s performance with different algorithm implementations was compared using mean average precision and inference time. The results of four different surface-mount components showed average precision scores of 97.3% and 97.7% for capacitor and resistor detection. The findings suggest that the system with the implemented YOLOv4-tiny algorithm on the Jetson Nano 4 GB microcomputer achieves a mean average precision score of 88.03% with an inference time of 56.4 ms and 87.98% mean average precision with 11.2 ms inference time on the Tesla P100 16 GB platform.

DepTSol: An Improved Deep-Learning- and Time-of-Flight-Based Real-Time Social Distance Monitoring Approach under Various Low-Light Conditions

Social distancing is an utmost reliable practice to minimise the spread of coronavirus disease (COVID-19). As the new variant of COVID-19 is emerging, healthcare organisations are concerned with controlling the death and infection rates. Different COVID-19 vaccines have been developed and administered worldwide. However, presently developed vaccine quantity is not sufficient to fulfil the needs of the world’s population. The precautionary measures still rely on personal preventive strategies. The sharp rise in infections has forced governments to reimpose restrictions. Governments are forcing people to maintain at least 6 feet (ft) of safe physical distance to stay safe. With summers, low-light conditions can become challenging. Especially in the cities of underdeveloped countries, where poor ventilated and congested homes cause people to gather in open spaces such as parks, streets, and markets. Besides this, in summer, large friends and family gatherings mostly take place at night. It is necessary to take precautionary measures to avoid more drastic results in such situations. To support the law and order bodies in maintaining social distancing using Social Internet of Things (SIoT), the world is considering automated systems. To address the identification of violations of a social distancing Standard Operating procedure (SOP) in low-light environments via smart, automated cyber-physical solutions, we propose an effective social distance monitoring approach named DepTSol. We propose a low-cost and easy-to-maintain motionless monocular time-of-flight (ToF) camera and deep-learning-based object detection algorithms for real-time social distance monitoring. The proposed approach detects people in low-light environments and calculates their distance in terms of pixels. We convert the predicted pixel distance into real-world units and compare it with the specified safety threshold value. The system highlights people violating the safe distance. The proposed technique is evaluated by COCO evaluation metrics and has achieved a good speed–accuracy trade-off with 51.2 frames per second (fps) and a 99.7% mean average precision (mAP) score. Besides the provision of an effective social distance monitoring approach, we perform a comparative analysis between one-stage object detectors and evaluate their performance in low-light environments. This evaluation will pave the way for researchers to study the field further and will enlighten the efficiency of deep-learning algorithms in timely responsive real-world applications.

A Novel classification framework for the Thirukkural for building an efficient search system

Thirukkural, a Tamil classic literature, which was written in 300 BCE is a didactic literature. Though Thirukkural comprises 1330 couplets which are organized into three sections and 133 chapters, in order to retrieve meaningful Thirukkural for a given query in search systems, a better organization of the Thirukkural is needed. This paper lays such a foundation by classifying the Thirukkural into ten new categories called superclasses that is helpful for building a better Information Retrieval (IR) system. The classifier is trained using Multinomial Naïve Bayes algorithm. Each superclass is further classified into two subcategories based on the didactic information. The proposed classification framework is evaluated using precision, recall and F-score metrics and achieved an overall F-score of 82.33% and a comparison analysis has been done with the Support Vector Machine, Logistic Regression and Random Forest algorithms. An IR system is built on top of the proposed system and the performance comparison has been done with the Google search and a locally built keyword search. The proposed classification framework has achieved a mean average precision score of 89%, whereas the Google search and keyword search have yielded 59% and 68% respectively.

Precise Agriculture: Effective Deep Learning Strategies to Detect Pest Insects

Pest insect monitoring and control is crucial to ensure a safe and profitable crop growth in all plantation types, as well as guarantee food quality and limited use of pesticides. We aim at extending traditional monitoring by means of traps, by involving the general public in reporting the presence of insects by using smartphones. This includes the largely unexplored problem of detecting insects in images that are taken in noncontrolled conditions. Furthermore, pest insects are, in many cases, extremely similar to other species that are harmless. Therefore, computer vision algorithms must not be fooled by these similar insects, not to raise unmotivated alarms. In this work, we study the capabilities of state-of-the-art (SoA) object detection models based on convolutional neural networks (CNN) for the task of detecting beetle-like pest insects on nonhomogeneous images taken outdoors by different sources. Moreover, we focus on disambiguating a pest insect from similar harmless species. We consider not only detection performance of different models, but also required computational resources. This study aims at providing a baseline model for this kind of tasks. Our results show the suitability of current SoA models for this application, highlighting how FasterRCNN with a MobileNetV3 backbone is a particularly good starting point for accuracy and inference execution latency. This combination provided a mean average precision score of 92.66% that can be considered qualitatively at least as good as the score obtained by other authors that adopted more specific models.

TRDet: Two-Stage Rotated Detection of Rural Buildings in Remote Sensing Images

Fast and accurate acquisition of the outline of rural buildings on remote sensing images is an efficient method to monitor illegal rural buildings. The traditional object detection method produces useless background information when detecting rural buildings; the semantic segmentation method cannot accurately segment the contours between buildings; the instance segmentation method cannot obtain regular building contours. The rotated object detection methods can effectively solve the problem that the traditional artificial intelligence method cannot accurately extract the outline of buildings. However, the rotated object detection methods are easy to lose location information of small objects in advanced feature maps and are sensitive to noise. To resolve these problems, this paper proposes a two-stage rotated object detection network for rural buildings (TRDet) by using a deep feature fusion network (DFF-Net) and a pixel attention module (PAM). Specifically, TRDet first fuses low-level location and high-level semantic information through the DFF-Net and then reduces the interference of noise information to the network through the PAM. The experimental results show that the mean average precession (mAP), precision, recall rate, and F1 score of the proposed TRDet are 83.57%, 91.11%, 86.5%, and 88.74%, respectively, which outperform the R2CNN model by 15%, 15.54%, 4.01%, and 9.87%. The results demonstrate that the TRDet can achieve better detection in small rural buildings and dense rural buildings.

Anomaly Detection in Fused Filament Fabrication Using Machine Learning.

Fused filament fabrication (FFF) has been widely used in various industries, and the adoption of technology is growing significantly. However, the FFF process has several disadvantages like inconsistent part quality and print repeatability. The occurrence of manufacturing-induced defects often leads to these shortcomings. This study aims to develop and implement an on-site monitoring system, which consists of a camera attached to the print head and the laptop that processes the video feed, for the extrusion-based 3D printers incorporating computer vision and object detection models to detect defects and make corrections in real-time. Image data from two classes of defects were collected to train the model. Various YOLO architectures were evaluated to study the ability to detect and classify printing anomalies such as under-extrusion and over-extrusion. Four of the trained models, YOLOv3 and YOLOv4 with "Tiny" variation, achieved a mean average precision score of >80% using the AP50 metric. Subsequently, two of the models (YOLOv3-Tiny 100 and 300 epochs) were optimized using Open Neural Network Exchange (ONNX) model conversion and ONNX Runtime to improve the inference speed. A classification accuracy rate of 89.8% and an inference speed of 70 frames per second were obtained. Before implementing the on-site monitoring system, a correction algorithm was developed to perform simple corrective actions based on defect classification. The G-codes of the corrective actions were sent to the printers during the printing process. This implementation successfully demonstrated real-time monitoring and autonomous correction during the FFF 3D printing process. This implementation will pave the way for an on-site monitoring and correction system through closed-loop feedback from other additive manufacturing (AM) processes.

Litchi detection in the field using an improved YOLOv3 model

Due to the illumination, complex background, and occlusion of the litchi fruits, the accurate detection of litchi in the field is extremely challenging. In order to solve the problem of the low recognition rate of litchi-picking robots in field conditions, this study was inspired by the ideas of ResNet and dense convolution and proposed an improved feature-extraction network model named “YOLOv3_Litchi”, combining dense connections and residuals for the detection of litchis. Firstly, based on the traditional YOLOv3 deep convolution neural network and regression detection, the idea of residuals was to be put into the feature-extraction network to effectively avoid the problem of decreasing detection accuracy due to the excessive depths of the network layers. Secondly, under the premise of a good receptive field and high detection accuracy, the large convolution kernel was replaced by a small convolution kernel in the shallow layer of the network, thereby effectively reducing the model parameters. Finally, the idea of feature pyramid was used to design the network to identify the small target litchi to ensure that the shallow features were not lost and simultaneously reduced the model parameters. Experimental results show that the improved YOLOv3_Litchi model achieved better results than the classic YOLOv3_DarkNet-53 model and the YOLOv3_Tiny model. The mean average precision (mAP) score was 97.07%, which was higher than the 95.18% mAP of the YOLOv3_DarkNet-53 model and the 94.48% mAP of the YOLOv3_Tiny model. The frame frequency was 58 fps, which was higher than 29 fps of the YOLOv3_DarkNet-53 model. Compared with the classic Faster R-CNN model with the feature-extraction network VGG16, the mAP was increased by 1%, and the FPS advantage was obvious. Compared with the classic single shot multibox detector (SSD) model, both the accuracy and the running efficiency were improved. The results show that the improved YOLOv3_Litchi model had stronger robustness, higher detection accuracy, and less computational complexity for the identification of litchi in the field conditions, which should be helpful for litchi orchard precision management. Keywords: deep learning; residual network; dense connection; feature pyramid network DOI: 10.25165/j.ijabe.20221502.6541 Citation: Peng H X, Xue C, Shao Y Y, Chen K Y, Liu H N, Xiong J T, et al. Litchi detection in the field using an improved YOLOv3 model. Int J Agric & Biol Eng, 2022; 15(2): 211–220.

Computer-Vision Based Object Detection and Recognition for Service Robot in Indoor Environment

The near future has been envisioned as a collaboration of humans with mobile robots to help in the day-to-day tasks. In this paper, we present a viable approach for a real-time computer vision based object detection and recognition for efficient indoor navigation of a mobile robot. The mobile robotic systems are utilized mainly for home assistance, emergency services and surveillance, in which critical action needs to be taken within a fraction of second or real-time. The object detection and recognition is enhanced with utilization of the proposed algorithm based on the modification of You Look Only Once (YOLO) algorithm, with lesser computational requirements and relatively smaller weight size of the network structure. The proposed computer-vision based algorithm has been compared with the other conventional object detection/recognition algorithms, in terms of mean Average Precision (mAP) score, mean inference time, weight size and false positive percentage. The presented framework also makes use of the result of efficient object detection/recognition, to aid the mobile robot navigate in an indoor environment with the utilization of the results produced by the proposed algorithm. The presented framework can be further utilized for a wide variety of applications involving indoor navigation robots for different services.

Induced Pluripotent Stem Cells Detection via Ensemble Yolo Network.

Induced pluripotent stem cells (iPSCs) have huge potential in regenerative medicine research and industrial applications. However, building automatic method without using cell staining technique for iPSCs identification is an important challenge. To improve the efficiency of producing iPSCs, we build an accurate and noninvasive iPSCs colonies detection method via ensemble Yolo network based on the self-collected bright-field microscopy images. Meanwhile, test-time augmentation (TTA) is leveraged to further improve the detection result of our iPSCs colonies detection method. Extensive experimental results on our dataset demonstrate that our method obtains quite favorable detection performance with the highest F1 score of 0.867 and the highest mean average precision score of 0.898, which outperforms most mainstream methods.

Tensor pooling-driven instance segmentation framework for baggage threat recognition

Automated systems designed for screening contraband items from the X-ray imagery are still facing difficulties with high clutter, concealment, and extreme occlusion. In this paper, we addressed this challenge using a novel multi-scale contour instance segmentation framework that effectively identifies the cluttered contraband data within the baggage X-ray scans. Unlike standard models that employ region-based or keypoint-based techniques to generate multiple boxes around objects, we propose to derive proposals according to the hierarchy of the regions defined by the contours. The proposed framework is rigorously validated on three public datasets, dubbed GDXray, SIXray, and OPIXray, where it outperforms the state-of-the-art methods by achieving the mean average precision score of 0.9779, 0.9614, and 0.8396, respectively. Furthermore, to the best of our knowledge, this is the first contour instance segmentation framework that leverages multi-scale information to recognize cluttered and concealed contraband data from the colored and grayscale security X-ray imagery.

Image Processing for Public Health Surveillance of Tobacco Point-of-Sale Advertising: Machine Learning-Based Methodology.

BackgroundWith a rapidly evolving tobacco retail environment, it is increasingly necessary to understand the point-of-sale (POS) advertising environment as part of tobacco surveillance and control. Advances in machine learning and image processing suggest the ability for more efficient and nuanced data capture than previously available.ObjectiveThe study aims to use machine learning algorithms to discover the presence of tobacco advertising in photographs of tobacco POS advertising and their location in the photograph.MethodsWe first collected images of the interiors of tobacco retailers in West Virginia and the District of Columbia during 2016 and 2018. The clearest photographs were selected and used to create a training and test data set. We then used a pretrained image classification network model, Inception V3, to discover the presence of tobacco logos and a unified object detection system, You Only Look Once V3, to identify logo locations.ResultsOur model was successful in identifying the presence of advertising within images, with a classification accuracy of over 75% for 8 of the 42 brands. Discovering the location of logos within a given photograph was more challenging because of the relatively small training data set, resulting in a mean average precision score of 0.72 and an intersection over union score of 0.62.ConclusionsOur research provides preliminary evidence for a novel methodological approach that tobacco researchers and other public health practitioners can apply in the collection and processing of data for tobacco or other POS surveillance efforts. The resulting surveillance information can inform policy adoption, implementation, and enforcement. Limitations notwithstanding, our analysis shows the promise of using machine learning as part of a suite of tools to understand the tobacco retail environment, make policy recommendations, and design public health interventions at the municipal or other jurisdictional scale.

P–247 Application of deep learning for automated measurement of key morphological features of human zygotes for IVF

Abstract Study question Can deep learning (DL) algorithms trained on time-lapse videos be used to detect and track the size and gender of pronuclei in developing human zygotes? Summary answer Our DL algorithm not only outperforms state-of-the-art models in detecting the pronuclei but can also accurately identify and track its gender and size over time. What is known already Recent researches have explored the use of DL to extract key morphological features of human embryos. Existing studies, however, focus either on blastocysts’ morphological measurements (Au et al. 2020) or on embryos’ general developmental stages classification (Gingold et al. 2018, Liu et al. 2019, Lau et al. 2019). So far, only one paper attempted to evaluate zygotes’ morphological components but stopped short of identifying the existence and location of their pronuclei (Leahy et al. 2020). We address this research gap by training a DL model that can detect, classify the gender, and quantify the size of zygotes’ pronuclei over time. Study design, size, duration A retrospective analysis using 91 fertilized oocytes from infertile patients undergoing IVF or ICSI treatment at Hanabusa Women’s Clinic between January 2011 and August 2019 was conducted. Each embryo was time-lapse monitored using Vitrolife which records an image every 15 minutes at 7 focal planes. For our study, we used videos of the first 1–2 days of the embryo from its 3 central focal planes, corresponding to 70–150 images per focal plane. Participants/materials, setting, methods All 273 timelapse videos were split into 30,387 grayscale still images at a 15-minute interval. Each image was checked and annotated by experienced embryologists where every pixel of the image was classified into 3 categories: male pronuclei, female pronuclei, and others. Images were converted into grayscale, resized into 500x500 pixels, and then fed into a neural network with the Mask R-CNN architecture and a ResNet101 backbone to produce a pronuclei instance segmentation model. Main results and the role of chance The 91 embryos were split into training (∼70% or 63 embryos) and validation (∼30% or 28 embryos). Our pronuclei model takes as input a single image and outputs a bounding box, mask, category, confidence score, and size measured in terms of pixel for each detected candidate. For prediction, we run the model on the 3 middle focal planes and merge candidates by using the one with the highest confidence score. We used the mean-average precision (mAP) score to evaluate our model’s ability to detect pronuclei and used the mean absolute percentage error (MAPE) between the actual size (as annotated by the embryologist) and the predicted one to check the model’s performance in tracking the pronuclei’s size. The mAP for detecting pronuclei, regardless of its gender, achieved by our model was 0.698, higher than the 0.680 value reported in the Leahy et al. paper (2020). Breakdown by gender, our model’s mAP for male and female pronuclei are 0.734 and 0.661 respectively. The overall MAPE for tracking pronuclei’s size is 21.8%. Breakdown by gender, our model’s MAPE for male and female pronuclei are 19.4% and 24.3% respectively. Limitations, reasons for caution Samples were collected from one clinic with videos recorded from one time-lapse system which can limit our results’ reproducibility. The accuracy of our DL model is also limited by the small number of embryos that we used. Wider implications of the findings: Even with a limited training dataset, our results indicate that we can accurately detect and track the gender and the size of zygotes’ pronuclei using time-lapse videos. In future models, we will increase our training dataset as well as include other time-lapse systems to improve our models’ accuracy and reproducibility. Trial registration number Not applicable

Deep learning applications based on SDSS photometric data: detection and classification of sources

ABSTRACT Most astronomical source classification algorithms based on photometric data struggle to classify sources as quasars, stars, and galaxies reliably. To achieve this goal and build a new Sloan Digital Sky Survey photometric catalogue in the future, we apply a deep learning source detection network built on YOLO v4 object detection framework to detect sources and design a new deep learning classification network named APSCnet (astronomy photometric source classification network) to classify sources. In addition, a photometric background image generation network is applied to generate background images in the process of data sets synthesis. Our detection network obtains a mean average precision score of 88.02 when IOU = 0.5. As for APSCnet, in a magnitude range with 14–25, we achieve a precision of 84.1 ${{\ \rm per\ cent}}$ at 93.2 ${{\ \rm per\ cent}}$ recall for quasars, a precision of 94.5 ${{\ \rm per\ cent}}$ at 84.6 ${{\ \rm per\ cent}}$ recall for stars, and a precision of 95.8 ${{\ \rm per\ cent}}$ at 95.1 ${{\ \rm per\ cent}}$ recall for galaxies; and in a magnitude range with less than 20, we achieve a precision of 96.6 ${{\ \rm per\ cent}}$ at 94.7${{\ \rm per\ cent}}$ recall for quasars, a precision of 95.7${{\ \rm per\ cent}}$ at 97.4${{\ \rm per\ cent}}$ recall for stars, and a precision of 98.9 ${{\ \rm per\ cent}}$ at 99.2 ${{\ \rm per\ cent}}$ recall for galaxies. We have proved the superiority of our algorithm in the classification of astronomical sources through comparative experiments between multiple sets of methods. In addition, we also analysed the impact of point spread function on the classification results. These technologies may be applied to data mining of the next generation sky surveys, such as LSST, WFIRST, and CSST etc.

Unsupervised anomaly instance segmentation for baggage threat recognition

Identifying potential threats concealed within the baggage is of prime concern for the security staff. Many researchers have developed frameworks that can automatically detect baggage threats from security X-ray scans. However, to the best of our knowledge, all of these frameworks require extensive training efforts on large-scale and well-annotated datasets, which are hard to procure in the real world, especially for the rarely seen contraband items. This paper presents a novel unsupervised anomaly instance segmentation framework that recognizes baggage threats, in X-ray scans, as anomalies without requiring any ground truth labels. Furthermore, thanks to its stylization capacity, the framework is trained only once, and at the inference stage, it detects and extracts contraband items regardless of their scanner specifications. Our one-staged approach initially learns to reconstruct normal baggage content via an encoder–decoder network utilizing a proposed stylization loss function. The model subsequently identifies the abnormal regions by analyzing the disparities within the original and the reconstructed scans. The anomalous regions are then clustered and post-processed to fit a bounding box for their localization. In addition, an optional classifier can also be appended with the proposed framework to recognize the categories of these extracted anomalies. A thorough evaluation of the proposed system on four public baggage X-ray datasets, without any re-training, demonstrates that it achieves competitive performance as compared to the conventional fully supervised methods (i.e., the mean average precision score of 0.7941 on SIXray, 0.8591 on GDXray, 0.7483 on OPIXray, and 0.5439 on COMPASS-XP dataset) while outperforming state-of-the-art semi-supervised and unsupervised baggage threat detection frameworks by 67.37%, 32.32%, 47.19%, and 45.81% in terms of F1 score across SIXray, GDXray, OPIXray, and COMPASS-XP datasets, respectively.

POS0036 AN ARTIFICIAL INTELLIGENCE MODEL IN RHEUMATOLOGY: INTERPRETATION OF THE SACROILIAC JOINT GRAPHY IN ANKYLOSING SPONDYLITIS

Background:There is an average of 8 years delay in the diagnosis of ankylosing spondylitis (AS). The most important danger of late diagnosis is that the disease can cause physical and functional disability (2). There is no specific diagnostic biomarker for AS. Sacroiliac joint (SIJ) radiography is frequently used in the diagnosis and follow-up of AS due to its easy accessibility and low cost. It can be classified as grade 0, 1, 2, 3, 4, and these classes may not be sharply separated from each other (3).Objectives:Interpretation of the SIJ radiography may differ from physician to physician. In fact, the same physician may interpret it differently at different times (3). We wanted to find a solution to the intraobserver disagreement problem with the artificial intelligence model.Methods:The SIJ radiography of 590 patients who applied to our center were divided into 3 categories as right and left, separately, grade 0, grade 1-2, grade 3-4, and an educational data set was prepared for the object recognition method. 488 images were augmented through noise from 490 images in the training data. 242 articular objects were trained for grade 0, 278 for grade 1-2, and 1426 for grade 2-3. The model was tested with 100 images for 36 joint objects for grade 0, 29 for grade 1-2, and 135 for grade 3-4 to create a computer vision-artificial intelligence model (image 1).Results:Training performance is 70% for grade 0, %63 for grade 1-2, %90 for grade 3-4 and test performance is %52 for grade 0, %24 for grade 1-2, %86 intersection over union (I/U:Intersection over Union is a form of measurement used to indicate the accuracy of an object detector.) for grade 3-4. The mean average precision (mAP) score of our object detection model is %65.9 for test data set (image 1). The estimation quality of the model can be affected by the distribution and number of each class.Conclusion:The experience of the x-ray technician, dose adjustment, and position differences due to patient compliance complicate the standardization of SIJ radiography and this may cause interobserver disagreement (3). Artificial intelligence models to be created with a larger and homogeneous data set in order to ensure objective standardization in the interpretation of the SIJ graph can help physicians.

Plant leaf classification and retrieval using multi‐scale shape descriptor

Plant leaf classification is a significant and challenging research problem in computer vision area. In this study, an original multi-scale shape descriptor is presented to perform leaf classification and retrieval. Firstly, a novel iterative rule is proposed as scales generation method, which is parameter free. Secondly, leaf contour points are represented by angle information which is calculated using their neighbour points under each scale. The angle information representation is invariant to image rotation, translation and scaling. More importantly, it can describe leaf in a hierarchical way by capturing leaf features from global to local variations. Then Fast Fourier Transform operation is applied to make the representation more compact and independent from starting point. Subsequently, for leaf retrieval the dissimilarity of each pair of leaf under each scale is computed using city block metric. And support vector machine is used as classifier for leaf classification. Finally, experiments and comparisons with multiple state-of-the-art approaches are performed. The classification accuracy was 96.85% and 93.56% respectively on Swedish and Flavia leaf datasets. The mean average precision score was 66.42%, 76.69% and 44.14% respectively on Flavia, Swedish and MEW2012 leaf datasets. The results demonstrate that the proposed method has excellent performance.

Video object detection with a convolutional regression tracker

Video object detection is a fundamental research task for scene understanding. Compared with object detection in images, object detection in videos has been less researched due to shortage of labelled video datasets. As frames in a video clip are highly correlated, a larger quantity of video labels are needed to have good data variation, which are not always available as the labels are much more expensive to attain. Regarding the above-mentioned problem, it is easy to train an image object detector, but not always possible to train a video object detector if there are insufficient video labels for certain classes. In order to deal with this problem and improve the performance of an image object detector for the classes without video labels, we propose to augment a well-trained image object detector with an efficient and effective class-agnostic convolutional regression tracker for the video object detection task. The tracker learns to track objects by reusing the features from the image object detector, which is a light-weighted increment to the detector, with only a slight speed drop for the video object detection task. The performance of our model is evaluated on the large-scale ImageNet VID dataset. Our strategy improves the mean average precision (mAP) score for the image object detector by around 5% and around 3% for the image object detector plus Seq-NMS post-processing.