Faster RCNN Inception V2 Research Articles

Comparison of Plant Detection Performance of CNN-based Single-stage and Two-stage Models for Precision Agriculture

The fact that arable land is not increasing in proportion to the ever-increasing population will increase the need for food in the coming years. For this reason, it is necessary to increase the yield of crops to make optimum use of arable land. One of the most important reasons for the decrease in yield and quality of crops is weeds. Herbicides are generally preferred for weed management. Due to deficiencies in herbicide application methods, only 0.015-6% of herbicides reach their target. The use of herbicides, which is an important part of the agricultural system, is an issue that needs to be emphasized, considering the risk of residue and environmental damage. In parallel with the rapid development of electronic and computer technologies, artificial intelligence applications have had the opportunity to develop. In this context, the use of artificial intelligence for plant detection in the subsystems of herbicide application machines will contribute to the development of precision agriculture techniques. In this study, the plant detection performances of single-stage and two-stage Convolutional Neural Network (CNN)-based deep learning (DL) models are evaluated. In this context, a dataset was created by taking images of Zea mays, Rhaponticum repens (L.) Hidalgo, and Chenopodium album L. plants in agricultural lands in Konya. With this dataset, the training of the models was carried out by the transfer learning method. The evaluation metrics of the trained models were calculated using the error matrix. In addition, training time and prediction time were used as quantitative metrics in the evaluation of the models. The plant detection performance, training time, and prediction time of the models were 85%, 8 h, 1.21 s for SSD MobileNet v2 and 99%, 22 h, 2.32 s for Faster R-CNN Inception v2, respectively. According to these results, Faster R-CNN Inception v2 is outperform in terms of accuracy. However, in cases where training time and prediction time are important, the SSD MobileNet v2 model can be trained with more data to increase its accuracy.

Traffic Lights Detection and Recognition with New Benchmark Datasets Using Deep Learning and TensorFlow Object Detection API

Today, traffic lights are widely used in places with high vehicle traffic. Especially in autonomous vehicles, fast and high accuracy detection and recognition of traffic lights are critical. Machine learning methods are generally used to do this. Deep learning models give more successful results than machine learning methods in detecting the exact location of traffic lights in different climatic conditions. In this study, Faster R-CNN Inception v2 deep learning model was trained and tested on two different datasets that we prepared and published publicly under variable traffic and climatic conditions in Turkey. Successful results were obtained with fewer data by using the Transfer Learning method with the help of TensorFlow Object Detection API in the training of the model. It has been shown that the datasets we have prepared can be developed considering the conditions in other countries and successful results will be obtained.

Face Mask Detection

Abstract: Various industries around the world were affected by the COVID-19 pandemic. Some sectors, like the development industry, have remained open despite the closures. The WHO has issued a warning for workers to wear a mask and avoid working in areas with high risks of infection. This paper developed a computing system that will automatically detect the presence of masks among workers on construction sites during the onset of the pandemic. It collected over a thousand images and added them to a database. The algorithm was trained and tested on various object detection models. It had been then ready to detect the presence of individuals using the Faster R-CNN Inception V2. The space between people was computed using the Euclidian distance. The model was then trained on various pictures and videos to spot the presence of masks.

Comparative Performance Analysis of Real-Time Methods for Cassava Phytoplasma Disease (CPD) Detection based on Deep Learning Neural Networks

Cassava Phytoplasma Disease (CPD) is a crop disease that reduces cassava output and quality. As a result, detection is essential in precision agriculture. On the greater area of the cassava field, manual identification of CPD illnesses takes more time and effort. Convolutional Neural Networks (CNN), a deep learning method, may be used to detect illnesses on leaves and other sections of cassava plants with greater accuracy. The approaches utilized in this study assisted in the identification of CPD by completing customized training/fine-tuning on three CNN models for object recognition: Faster R-CNN, SSD Mobilenet v2, and YOLO v4. The Faster R-CNN inception v2 has a 95 percent training accuracy, SSD Mobilenet v2 has a 73 percent training accuracy, and YOLOv4 has an 85 percent training accuracy, according to the data. Finally, the study found that the YOLOv4 outperforms the Faster R-CNN inception v2 and SSD MobileNet v2 in terms of image computing capacity. However, Faster R-CNN inception v2 performs the best compared to the two other models in terms of accuracy. Hence, these two models can be used depending on the purpose of CPD detection. However, since CPD detection is the main purpose of this study, the Faster R-CNN model is recommended for adoption to detect CPD in a real-time environment. Keywords — cassava phytoplasma disease, convolutional neural networks, faster R-CNN, image processing, precision agricultur

An Automatic System to Monitor the Physical Distance and Face Mask Wearing of Construction Workers in COVID-19 Pandemic.

The COVID-19 pandemic has caused many shutdowns in different industries around the world. Sectors such as infrastructure construction and maintenance projects have not been suspended due to their significant effect on people’s routine life. In such projects, workers work close together that makes a high risk of infection. The World Health Organization recommends wearing a face mask and practicing physical distancing to mitigate the virus’s spread. In this paper, we developed a computer vision system to automatically detect the violation of face mask wearing and physical distancing among construction workers to assure their safety on infrastructure projects during the pandemic. For the face mask detection, we collected and annotated 1000 images, including different types of face mask wearing, and added them to a pre-existing face mask dataset to develop a dataset of 1853 images and increased the dataset to 3300 images by data augmentation. Then, we trained and tested multiple Tensorflow state-of-the-art object detection models on the face mask dataset and chose the Faster R-CNN Inception ResNet V2 network that yielded the accuracy of 99.8%. For physical distance detection, we employed the Faster R-CNN Inception V2 to detect people. A transformation matrix was used to eliminate the camera angle’s effect on the object distances on the image. The Euclidian distance used the pixels of the transformed image to compute the actual distance between people. A threshold of six feet was considered to capture physical distance violation. We also used transfer learning for training the model. The final model was applied on four videos of road maintenance projects in Houston, TX, that effectively detected the face mask and physical distance. We recommend that construction owners use the proposed system to enhance construction workers’ safety in the pandemic situation.

Diagnostic charting of panoramic radiography using deep-learning artificial intelligence system.

The goal of this study was to develop and evaluate the performance of a new deep-learning (DL) artificial intelligence (AI) model for diagnostic charting in panoramic radiography. One thousand eighty-four anonymous dental panoramic radiographs were labeled by two dento-maxillofacial radiologists for ten different dental situations: crown, pontic, root-canal treated tooth, implant, implant-supported crown, impacted tooth, residual root, filling, caries, and dental calculus. AI Model CranioCatch, developed in Eskişehir, Turkey and based on a deep CNN method, was proposed to be evaluated. A Faster R-CNN Inception v2 (COCO) model implemented with the TensorFlow library was used for model development. The assessment of AI model performance was evaluated with sensitivity, precision, and F1 scores. When the performance of the proposed AI model for detecting dental conditions in panoramic radiographs was evaluated, the best sensitivity values were obtained from the crown, implant, and impacted tooth as 0.9674, 0.9615, and 0.9658, respectively. The worst sensitivity values were obtained from the pontic, caries, and dental calculus, as 0.7738, 0.3026, and 0.0934, respectively. The best precision values were obtained from pontic, implant, implant-supported crown as 0.8783, 0.9259, and 0.8947, respectively. The worst precision values were obtained from residual root, caries, and dental calculus, as 0.6764, 0.5096, and 0.1923, respectively. The most successful F1 Scores were obtained from the implant, crown, and implant-supported crown, as 0.9433, 0.9122, and 0.8947, respectively. The proposed AI model has promising results at detecting dental conditions in panoramic radiographs, except for caries and dental calculus. Thanks to the improvement of AI models in all areas of dental radiology, we predict that they will help physicians in panoramic diagnosis and treatment planning, as well as digital-based student education, especially during the pandemic period.

An artif\u0131cial \u0131ntelligence approach to automatic tooth detection and numbering in panoramic radiographs

BackgroundPanoramic radiography is an imaging method for displaying maxillary and mandibular teeth together with their supporting structures. Panoramic radiography is frequently used in dental imaging due to its relatively low radiation dose, short imaging time, and low burden to the patient. We verified the diagnostic performance of an artificial intelligence (AI) system based on a deep convolutional neural network method to detect and number teeth on panoramic radiographs.MethodsThe data set included 2482 anonymized panoramic radiographs from adults from the archive of Eskisehir Osmangazi University, Faculty of Dentistry, Department of Oral and Maxillofacial Radiology. A Faster R-CNN Inception v2 model was used to develop an AI algorithm (CranioCatch, Eskisehir, Turkey) to automatically detect and number teeth on panoramic radiographs. Human observation and AI methods were compared on a test data set consisting of 249 panoramic radiographs. True positive, false positive, and false negative rates were calculated for each quadrant of the jaws. The sensitivity, precision, and F-measure values were estimated using a confusion matrix.ResultsThe total numbers of true positive, false positive, and false negative results were 6940, 250, and 320 for all quadrants, respectively. Consequently, the estimated sensitivity, precision, and F-measure were 0.9559, 0.9652, and 0.9606, respectively.ConclusionsThe deep convolutional neural network system was successful in detecting and numbering teeth. Clinicians can use AI systems to detect and number teeth on panoramic radiographs, which may eventually replace evaluation by human observers and support decision making.

Deep Learning as innovation vector in the provisioning of new household electricity connections / Deep Learning como vetor de inovação no atendimento a solicitações de novas ligações elétricas residenciais

This paper presents a Deep Learning model to visually identify the energy meter box (EMB) in households able to be connected to the power grid. The model is image-based and aims at supporting an electric utility to provide new household electricity connections. The EMB identification is carried out by computer vision and occurs during the client's digital request journey, which is held on via WhatsApp. In such a journey the client interacts with a virtual assistant and sends a photo of the box that must already be installed. The model, developed within an R&D project for digital transformation, can identify the EMB’s presence and classify whether its facilities is single- or three-phase, streamlining the subsequent work of the technical team in the field. The paper describes the pipeline adopted for machine learning, considering techniques for the dataset generation, such as data augmentation technique used to expand the image bank for training and testing, and the neural networks used for training. Among the evaluated networks, “Faster RCNN Inception V2” generated the best results when analyzing the metrics and test inferences. The model resulting from the training stage was made available on a virtual machine – the IBM Virtual Private Server.

Indonesian License Plate Detection and Identification Using Deep Learning

Abstract— License plate is the unique identity of the vehicle, which serves as proof of the legitimacy of the operation of the vehicle in the form of a plate or other material with certain specifications issued by the police and contains the area code, registration number and validity period and installed on the vehicle. License plates are often used in automated parking systems and vehicle identification in case of traffic violations. So, it is necessary to build a system for detection and identification of license plates. The proposed license plate detection and identification system is divided into three main processes, namely license plate detection, character segmentation, and character recognition. The detection process uses transfer learning techniques using Faster R-CNN Inception V2. The segmentation process uses traditional computer vision with morphological operations and contours extraction. Then the character recognition process uses the MobileNet V2 transfer learning technique as an architecture for character classification. The recognition accuracy compared between MobileNet V2 and TesseractOCR shows that MobileNet V2 is superior with an accuracy rate of 96%, while Tesseract-OCR has a poor accuracy of 59%. Keywords— Deep Learning, Convolutional Neural Network, License Plate Detection, Character Segmentation, Character Recognition

Artificial intelligence system for automatic deciduous tooth detection and numbering in panoramic radiographs.

This study evaluated the use of a deep-learning approach for automated detection and numbering of deciduous teeth in children as depicted on panoramic radiographs. An artificial intelligence (AI) algorithm (CranioCatch, Eskisehir-Turkey) using Faster R-CNN Inception v2 (COCO) models were developed to automatically detect and number deciduous teeth as seen on pediatric panoramic radiographs. The algorithm was trained and tested on a total of 421 panoramic images. System performance was assessed using a confusion matrix. The AI system was successful in detecting and numbering the deciduous teeth of children as depicted on panoramic radiographs. The sensitivity and precision rates were high. The estimated sensitivity, precision, and F1 score were 0.9804, 0.9571, and 0.9686, respectively. Deep-learning-based AI models are a promising tool for the automated charting of panoramic dental radiographs from children. In addition to serving as a time-saving measure and an aid to clinicians, AI plays a valuable role in forensic identification.

NaturalAE: Natural and robust physical adversarial examples for object detectors

Recently, many studies show that deep neural networks (DNNs) are susceptible to adversarial examples, which are generated by adding imperceptible perturbations to the input of DNN. However, in order to convince that adversarial examples are real threats in real physical world, it is necessary to study and evaluate the adversarial examples in real-world scenarios. In this paper, we propose a natural and robust physical adversarial example attack method targeting object detectors under real-world conditions, which is more challenging than targeting image classifiers. The generated adversarial examples are robust to various physical constraints and visually look similar to the original images, thus these adversarial examples are natural to humans and will not cause any suspicions. First, to ensure the robustness of the adversarial examples in real-world conditions, the proposed method exploits different image transformation functions (Distance, Angle, Illumination and Photographing), to simulate various physical changes during the iterative optimization of the adversarial examples generation. Second, to construct natural adversarial examples, the proposed method uses an adaptive mask to constrain the area and intensities of the added perturbations, and utilizes the real-world perturbation score (RPS) to make the perturbations be similar to those real noises in physical world. Compared with existing studies, our generated adversarial examples can achieve a high success rate with less conspicuous perturbations. Experimental results demonstrate that, the generated adversarial examples are robust under various indoor and outdoor physical conditions, including different distances, angles, illuminations, and photographing. Specifically, the attack success rate of generated adversarial examples indoors and outdoors is high up to 73.33% and 82.22%, respectively. Meanwhile, the proposed method ensures the naturalness of the generated adversarial example, and the size of added perturbations is as low as 29361.86, which is much smaller than the perturbations in the existing works (95381.14 at the highest). Further, the proposed physical adversarial attack method can be transferred from the white-box models to other object detection models. The attack success rate of the adversarial examples (generated targeting Faster R-CNN Inception v2) is high up to 57.78% on the SSD models, while the success rate of adversarial example (generated targeting YOLO v2) on SSD models reaches 77.78%. This paper reveals that physical adversarial example attacks are real threats in the real-world conditions, and can hopefully provide guidance for designing robust object detectors and image classifiers.

FINE-TUNING DEEP LEARNING MODELS FOR PEDESTRIAN DETECTION

Object detection in high resolution images is a new challenge that the remote sensing community is facing thanks to introduction of unmanned aerial vehicles and monitoring cameras. One of the interests is to detect and trace persons in the images. Different from general objects, pedestrians can have different poses and are undergoing constant morphological changes while moving, this task needs an intelligent solution. Fine-tuning has woken up great interest among researchers due to its relevance for retraining convolutional networks for many and interesting applications. For object classification, detection, and segmentation fine-tuned models have shown state-of-the-art performance. In the present work, we evaluate the performance of fine-tuned models with a variation of training data by comparing Faster Region-based Convolutional Neural Network (Faster R-CNN) Inception v2, Single Shot MultiBox Detector (SSD) Inception v2, and SSD Mobilenet v2. To achieve the goal, the effect of varying training data on performance metrics such as accuracy, precision, F1-score, and recall are taken into account. After testing the detectors, it was identified that the precision and recall are more sensitive on the variation of the amount of training data. Under five variation of the amount of training data, we observe that the proportion of 60%-80% consistently achieve highly comparable performance, whereas in all variation of training data Faster R-CNN Inception v2 outperforms SSD Inception v2 and SSD Mobilenet v2 in evaluated metrics, but the SSD converges relatively quickly during the training phase. Overall, partitioning 80% of total data for fine-tuning trained models produces efficient detectors even with only 700 data samples.

Efficient inception V2 based deep convolutional neural network for real‐time hand action recognition

The most effective and accurate deep convolutional neural network (faster region-based convolutional neural network (Faster R-CNN) Inception V2 model, single shot detector (SSD) Inception V2 model) based architectures for real-time hand gesture recognition is proposed. The proposed models are tested on standard data sets (NUS hand posture data set-II, Senz-3D) and custom-developed (MITI hand data set (MITI-HD)) data set. The performance metrics are analysed for intersection over union (IoU) ranges between 0.5 and 0.95. IoU value of 0.5 resulted in higher precision compared to other IoU values considered (0.5:0.95, 0.75). It is observed that the Faster R-CNN Inception V2 model resulted in higher precision (0.990 for APall, IoU = 0.5) compared to SSD Inception V2 model (0.984 for all) for MITI-HD 160. The computation time of Faster R-CNN Inception V2 is higher compared to SSD Inception V2 model and also resulted in less number of mispredictions. Increasing the size of samples (MITI-HD 300) resulted in improvement of APall = 0.991. Improvement in large (APlarge) and medium (APmedium) size detections are not significant when compared to small (APsmall) detections. It is concluded that the Faster R-CNN Inception V2 model is highly suitable for real-time hand gesture recognition system under unconstrained environments.

Bridge Girder Crack Assessment Using Faster RCNN Inception V2 and Infrared Thermography

Manual inspections of infrastructures such as highway bridge, pavement, dam, and multistoried garage ceiling are time consuming, sometimes can be life threatening, and costly. An automated computerized system can reduce time, faulty inspection, and cost of inspection. In this study, we developed a computer model using deep learning Convolution Neural Network (CNN), which can be used to automatically detect the crack and non-crack type structure. The goal of this research is to allow application of state-of-the-art deep neural network and Unmanned Aerial Vehicle (UAV) technologies for highway bridge girder inspection. As a pilot study of implementing deep learning in Bridge Girder, we study the recognition, length, and location of crack in the structure of the UTC campus old garage concrete ceiling slab. A total of 2086 images of crack and non-crack were taken from UTC Old Library parking garage ceiling using handheld mobile phone and drone. After training the model shows 98% accuracy with crack and non-crack types of structures.

An Automated Training of Deep Learning Networks by 3D Virtual Models for Object Recognition

Small series production with a high level of variability is not suitable for full automation. So, a manual assembly process must be used, which can be improved by cooperative robots and assisted by augmented reality devices. The assisted assembly process needs reliable object recognition implementation. Currently used technologies with markers do not work reliably with objects without distinctive texture, for example, screws, nuts, and washers (single colored parts). The methodology presented in the paper introduces a new approach to object detection using deep learning networks trained remotely by 3D virtual models. Remote web application generates training input datasets from virtual 3D models. This new approach was evaluated by two different neural network models (Faster RCNN Inception v2 with SSD, MobileNet V2 with SSD). The main advantage of this approach is the very fast preparation of the 2D sample training dataset from virtual 3D models. The whole process can run in Cloud. The experiments were conducted with standard parts (nuts, screws, washers) and the recognition precision achieved was comparable with training by real samples. The learned models were tested by two different embedded devices with an Android operating system: Virtual Reality (VR) glasses, Cardboard (Samsung S7), and Augmented Reality (AR) smart glasses (Epson Moverio M350). The recognition processing delays of the learned models running in embedded devices based on an ARM processor and standard x86 processing unit were also tested for performance comparison.

A DEEP LEARNING STUDY OF EXTRACTING NAVIGATION AREA FROM CAD BLUEPRINTS

Abstract. Deep learning technology is a cutting edge topic of AI region, and draws more attention from photogrammetry and remote sensing society. In this study, we strive to combine deep learning with CAD designs to extract navigation area (room). To this, we mark more than 200 2D building blueprint in CAD forms to construct the learning set to train object detection model based on TensorFlow. This model is the faster R-CNN inception v2 model from COCO dataset. The test and result section is composed of three parts: First part demonstrates the model performance on learning dataset; second part applies the generated model to extract rooms from untrained raw CAD blueprints; Third part covers the comparison between deep learning extracted result and geometric based algorithm extracted result. Test result shows that the deep learning approach could achieve higher accuracy than geometric approach under regular shape situations. In conclusion, we have proposed a well-trained deep learning model that could be utilized to construct a schema of the navigation area for 2D CAD blueprints.