Tiny YOLOv3 Research Articles

Deep learning neural network-assisted badminton movement recognition and physical fitness training optimization

This work aims to solve the problem of low accuracy in recognizing the trajectory of badminton movement. This work focuses on the visual system in badminton robots and conducts side detection and tracking of flying badminton in two-dimensional image plane video streams. Then, the cropped video images are input into a convolutional neural network frame by frame. By adding an attention mechanism, it helps identify the badminton movement trajectory. Finally, to address the detection challenge of flying badminton as a small target in video streams, the deep learning one-stage detection network, Tiny YOLOv2, is improved from both the loss function and network structure perspectives. Moreover, it is combined with the Unscented Kalman Filter algorithm to predict the trajectory of badminton movement. Simulation results show that the improved algorithm performs excellently in tracking and predicting badminton trajectories compared with the existing algorithms. The average accuracy of the proposed method for tracking badminton trajectories is 91.40 %, and the recall rate is 84.60 %. The average precision, recall, and frame rate of the measured trajectories in four simple and complex scenarios of badminton flight video streams are 96.7 %, 95.7 %, and 29.2 frames/second, respectively. They are all superior to other classic algorithms. It is evident that the proposed method can provide powerful support for badminton trajectory recognition and help improve the accuracy of badminton movement recognition.

Hardware acceleration of Tiny YOLO deep neural networks for sign language recognition: A comprehensive performance analysis

In this paper, we benchmark two automation frameworks, Vitis AI and FINN, for sign language recognition on a Field Programmable Gate Array (FPGA). We conducted an in-depth exploration of both frameworks using Tiny YOLOv2 networks by varying design parameters such as precision, parallelism ratio, etc. Further, a fair baseline comparison is made based on accuracy, speed, and hardware resources. Experimental findings demonstrate that the Vitis AI outperforms the FINN framework and traditional GPU and CPU platforms by achieving significant improvements of 1.08x, 1.7x, and 2.9x in terms of latency. Leveraging Vitis AI, our system achieved a detection speed of 32.7 frames per second (FPS) on the Kria KV260 FPGA with a power consumption rate of 5.6 W and an impressive mean Average Precision (mAP) score of 61.2% on the Hindi Indian Sign Language (ISL) dataset.

Comparative analysis of computer vision algorithms for the real-time detection of digital dermatitis in dairy cows

Digital dermatitis (DD) is a bovine claw disease responsible for ulcerative lesions on the planar aspect of the hoof. DD is associated with massive herd outbreaks of lameness and influences cattle welfare and production. Early detection of DD can lead to prompt treatment and decrease lameness. Computer vision (CV) provides a unique opportunity to improve early detection. The study aims to train and compare applications for the real-time detection of DD in dairy cows. Eight CV models were trained for detection and scoring, compared using performance metrics and inference time, and the best model was automated for real-time detection using images and video. Images were collected from commercial dairy farms while facing the interdigital space on the plantar surface of the foot. Images were scored for M-stages of DD by a trained investigator using the M-stage DD classification system with distinct labels for hyperkeratosis (H) and proliferations (P). Two sets of images were compiled: the first dataset (Dataset 1) containing 1,177 M0/M4H and 1,050 M2/M2P images and the second dataset (Dataset 2) containing 240 M0, 17 M2, 51 M2P, 114 M4H, and 108 M4P images. Models were trained to detect and score DD lesions and compared for precision, recall, and mean average precision (mAP) in addition to inference time in frame per second (FPS). Seven of the nine CV models performed well compared to the ground truth of labeled images using Dataset 1. The six models, Faster R-CNN, Cascade R-CNN, YOLOv3, Tiny YOLOv3, YOLOv4, Tiny YOLOv4, and YOLOv5s achieved an mAP between 0.964 and 0.998, whereas the other two models, SSD and SSD Lite, yielded an mAP of 0.371 and 0.387 respectively. Overall, YOLOv4, Tiny YOLOv4, and YOLOv5s outperformed all other models with almost perfect precision, perfect recall, and a higher mAP. Tiny YOLOv4 outperformed all other models with respect to inference time at 333 FPS, followed by YOLOv5s at 133 FPS and YOLOv4 at 65 FPS. YOLOv4 and Tiny YOLOv4 performed better than YOLOv5s compared to the ground truth using Dataset 2. YOLOv4 and Tiny YOLOv4 yielded a similar mAP of 0.896 and 0.895, respectively. However, Tiny YOLOv4 achieved both higher precision and recall compared to YOLOv4. Finally, Tiny YOLOv4 was able to detect DD lesions on a commercial dairy farm with high performance and speed. The proposed CV tool can be used for early detection and prompt treatment of DD in dairy cows. This result is a step towards applying CV algorithms to veterinary medicine and implementing real-time DD detection on dairy farms.

A novel optimized tiny YOLOv3 algorithm for the identification of objects in the lawn environment

Based on the problem of insufficient accuracy of the original tiny YOLOv3 algorithm for object detection in a lawn environment, an Optimized tiny YOLOv3 algorithm with less computation and higher accuracy is proposed. Three reasons affect the accuracy of the original tiny YOLOv3 algorithm for detecting objects in a lawn environment. First, the backbone of the original algorithm is composed of a stack of a single convolutional layer and a max-pooling layer, which results in insufficient ability to extract feature information of objects. An enhancement module is proposed to enhance the feature extraction capability of the shallow layers of the network. Second, the information of the shallow convolutional layers of the backbone is not fully used, which results in insufficient detection capability for small objects. Third, the deep part of the backbone uses a convolutional layer with an excessive number of channels, which results in a large amount of computation. A multi-resolution fusion module is proposed to enhance the information interaction capability of the deep and shallow layers of the network, and reduce the computation. To verify the accuracy of this Optimized tiny YOLOv3 algorithm, the algorithm was tested on the dataset containing trunk, spherical tree and person, and compared with the current research. The results show that the algorithm proposed in this paper improves the detection accuracy while reducing the calculation.

Efficient Object Detection Acceleration Methods for Autonomous-driving Embedded Platforms

Object detection in autonomous vehicles is typically operated in an embedded system to reduce power consumption. The use of an object detection algorithm with high accuracy and real-time detection speed in the embedded systems is essential for ensuring safe driving. This study proposes a parallel processing method for GPU and CPU operations to enhance the detection speed of the model. In addition, this study proposes data augmentation and image resize techniques that consider the camera input size of autonomous driving, which increases the accuracy significantly while improving the detection speed. The application of these proposed schemes to a baseline algorithm, tiny Gaussian YOLOv3, improves the mean average precision by 1.14 percent points (pp) for the Berkeley Deep Drive (BDD) dataset and 1.34 pp for the KITTI dataset compared to the baseline. Furthermore, in the NVIDIA Jetson AGX Xavier, which is an embedded platform for autonomous driving, the proposed algorithm improves the detection speed by 22.54 % for the BDD, and 24.67 % for the KITTI compared to the baseline, thereby enabling high-speed real-time detection on both datasets.

Application of Artificial Intelligence SSD MobileNet and Tiny YOLOv2 for Food Recipe Search

Recipes are guides to making something together with notes on ingredients and their amount. To be able to make food, of course, the cook must prepare the ingredients in advance to be processed into ready-to-eat dishes. Often people have a lot of food ingredients but don't know how to process them. Cooking without seriousness, of course, some people fail when making a dish. Back then, people depends on recipe that was passed down from generation to generation. Now, the digital world is growing rapidly. Anything can be done with increasingly modern technology. Everything needed is accessible with today's technology. Everything is so easy, including the matter of food. Even so, in this digital era, people use smartphones but still cannot use them properly. Many of them use search engines so they need to sort out which are real recipes and which are just random recipes. The purpose of this study is to help people find recipes by taking photos of food ingredients and then finding out what can be made from these ingredients. This technique uses Artificial Intelligence (AI) with MobileNet and Tiny YOLOv2 SSD modules. The design uses the Unified Modeling Language (UML). The study used experimental methods to test the accuracy of the AI used. Data collection will be utilizing a literature study. This research uses agile for system development. Test results show that the SSD MobileNet model has a guessing accuracy of 77%, while Tiny YOLOv2 is 81%. The guessing accuracy might get higher if good camera quality is used

Impact of Online Education and Sentiment Analysis from Twitter Data using Topic Modeling Algorithms

During a pandemic, all industries suffer greatly, and every sector of the world suffers in some way, including the education sector. Internet expressions reflect users' feelings about a product or service. The polarity of information in source data toward a subject under investigation is determined by sentiment analysis processes. The goal of this study is to examine social media expressions about online teaching and learning, as online education will become a part of everyday life in the future. We collected data from Twitter using keywords related to online education and Google form from engineering undergraduate students for prototype implementation. This analysis will assist teachers, parents, and the student community in understanding the benefits and drawbacks of the education industry, allowing for further improvement in educational outcomes. We used aspect-based sentiment analysis and topic modeling to determine sentiment polarity and important topics for education sector stakeholders. To begin, we used TextBlob Python package to determine sentiment polarity, and Bag of Words, LDA and LSA model for discovering topics. After modeling topics from the collected data, topic Coherence is used to assess the degree of semantic similarity between high scoring words in the topic. The word cloud and LDAvis are used to visualize data. The experimental results are promising and it will assist education stakeholders in addressing the concerns that have been identified as social media expressions to work on. Since the boom in science and technology, humans have been trying to invent machines that could reduce their efforts in day to day activities. In this paper, we develop a personal assistant robot that could pick up objects and return it to the user. The robot is controlled using an android application in mobile phones. The robot can listen to user’s command and then respond in the best way possible. The user can command the robot to move to given location, capture images and pick objects. The robot is equipped with ultrasonic sensor and web camera that helps it to move to different location effectively. It is also equipped with sleds that play important role in object picking process. The robot uses a tiny YOLOv3 model which is rigorously trained on several images of the object. There are some possible improvements that can be achieved which could help this robot to be used in several other fields as well.

Fast identification model for coal and gangue based on the improved tiny YOLO v3

Fast identification model for coal and gangue based on the improved tiny YOLO v3

Text-Based Traffic Panels Detection using the Tiny YOLOv3 Algorithm

Text-Based Traffic Panels Detection using the Tiny YOLOv3 Algorithm by Saba Kheirinejad 1,* , Noushin Riahi 2, Reza Azmi 2 1 Center for Ubiquitous Computing, University of Oulu, Oulu, Finland 2 Department of Computer Engineering, Faculty of Engineering, Alzahra University, Tehran, Iran * Author to whom correspondence should be addressed. Journal of Engineering Research and

Energy-Efficient Dataflow Scheduling of CNN Applications for Vector-SIMD DSP

Dataflow-scheduling techniques for convolutional neural networks (CNNs) are extensively studied to minimize the off-chip memory access. However, the efficiencies of the previously proposed techniques are limited because their optimizations only consider the general hardware such as FPGA and GPU. To overcome this limitation, this paper proposes dataflow scheduling for vector-SIMD DSP to minimize the energy consumption for the off-chip memory access. First, the proposed technique attempts to group as many given layers as possible. For grouping the layers, the tiles in different layers are executed in sequence without the off-chip memory access except the first and the last layers in the group. The length of the grouped layers is determined with regard to the minimization of the energy consumption of off-chip memory by estimating the proposed energy model of the off-chip memory. However, grouping the layers results in the additional computation. To minimize this overhead, this paper solves the optimization problem for in the grouped layers. Second, for layers that cannot be grouped, the tiling along the W-axis is not considered, to maximize the size of the overlapped data in consecutive tiles. Consequently, the reuse of the overlapped data in the on-chip buffer is maximized, thereby reducing the energy consumption by the off-chip memory. For evaluation, a cycle-accurate simulation environment is established to measure the energy consumption of the off-chip memory by tracing the data between a vector-SIMD DSP and an off-chip memory. The experimental results show that compared with the baseline tiling and scheduling techniques, the proposed technique reduces the energy consumption by an average of 51% for CNN applications such as Tiny YOLOv2, MobileNetv1, VDSR.

Personal Assistant Robot

Since the boom in science and technology, humans have been trying to invent machines that could reduce their efforts in day to day activities. In this paper, we develop a personal assistant robot that could pick up objects and return it to the user. The robot is controlled using an android application in mobile phones. The robot can listen to user’s command and then respond in the best way possible. The user can command the robot to move to given location, capture images and pick objects. The robot is equipped with ultrasonic sensor and web camera that helps it to move to different location effectively. It is also equipped with sleds that play important role in object picking process. The robot uses a tiny YOLOv3 model which is rigorously trained on several images of the object. There are some possible improvements that can be achieved which could help this robot to be used in several other fields as well.

Unsupervised Anomaly Detection in Printed Circuit Boards through Student–Teacher Feature Pyramid Matching

Deep learning methods are currently used in industries to improve the efficiency and quality of the product. Detecting defects on printed circuit boards (PCBs) is a challenging task and is usually solved by automated visual inspection, automated optical inspection, manual inspection, and supervised learning methods, such as you only look once (YOLO) of tiny YOLO, YOLOv2, YOLOv3, YOLOv4, and YOLOv5. Previously described methods for defect detection in PCBs require large numbers of labeled images, which is computationally expensive in training and requires a great deal of human effort to label the data. This paper introduces a new unsupervised learning method for the detection of defects in PCB using student–teacher feature pyramid matching as a pre-trained image classification model used to learn the distribution of images without anomalies. Hence, we extracted the knowledge into a student network which had same architecture as the teacher network. This one-step transfer retains key clues as much as possible. In addition, we incorporated a multi-scale feature matching strategy into the framework. A mixture of multi-level knowledge from the features pyramid passes through a better supervision, known as hierarchical feature alignment, which allows the student network to receive it, thereby allowing for the detection of various sizes of anomalies. A scoring function reflects the probability of the occurrence of anomalies. This framework helped us to achieve accurate anomaly detection. Apart from accuracy, its inference speed also reached around 100 frames per second.

Deep learning computer vision for robotic disassembly and servicing applications

Fastener detection is a necessary step for computer vision (CV) based robotic disassembly and servicing applications. Deep learning (DL) provides a robust approach for creating CV models capable of generalizing to diverse visual environments. Such DL CV systems rely on tuning input resolution and mini-batch size parameters to fit the needs of the detection application. This paper provides a method for determining the optimal compromise between input resolution and mini-batch size to determine the highest performance for cross-recessed screw (CRS) detection while utilizing maximum graphics processing unit resources. The Tiny-You Only Look Once v2 (Tiny-YOLO v2) DL object detection system was chosen to evaluate this method. Tiny-YOLO v2 was employed to solve the specialized task of detecting CRS which are highly common in electronic devices. The method used in this paper for CRS detection is meant to lay the ground-work for multi-class fastener detection, as the method is not dependent on the type or number of object classes. An original dataset of 900 images of 12.3 MPx resolution was manually collected and annotated for training. Three additional distinct datasets of 90 images each were manually collected and annotated for testing. It was found an input resolution of 1664 x 1664 pixels paired with a mini-batch size of 16 yielded the highest average precision (AP) among the seven models tested for all three testing datasets. This model scored an AP of 92.60% on the first testing dataset, 99.20% on the second testing dataset, and 98.39% on the third testing dataset.

Design and implementation of fish freshness detection algorithm using deep learning

Organoleptic assessment of fresh fish includes specifications for the quality of the eyes, gills, mucus, odor, texture and flesh (color and appearance). However, not everyone has knowledge about it. This research uses the tiny yolov2 to facilitate the determination of fish freshness levels (good quality, medium quality, poor quality) correctly and fast. There are a few stages in this research, included organoleptic test accompanied by taking fish eye image dataset every hour, processing organoleptic test data labeling, training, and validation. There are three types of fish used, consists of Rastrelliger, Euthynnus affinis, and Chanos chanos. Detection of fish freshness level for three species was successfully carried out with the result of average precision is 72.9%, average recall is 57.5%, and accuracy is 57.5%. The factors that affect the prediction results in this study is the collection of datasets before the training process is carried out consisting of fish samples obtained from traditional markets, which are considered inadequate so that it affects the organoleptic test process itself, the organoleptic test that was carried out as a reference for image sorting was considered inaccurate because it used less than 30 untrained panelists and dataset imbalance.

Lightweight target detection model for embedded platform

The detection speed of target detection algorithm depends on the performance of computer equipment. Aiming at the problems of slow detection speed and difficult trade-off between detection accuracy and detection speed when the target detection model is used in embedded devices, a lightweight target detection model based on the improved Tiny YOLO-V3 is proposed. Firstly, we analyze the time consumption of each layer structure in the convolutional neural network, and do a lot of experiments and tests. Then, we compress the time-consuming structure substantially. Secondly, we propose the segmentation and fusion module to improve the detection accuracy. Finally, we use the remote sensing data set of Wuhan University for experiments, and the experimental results show that compared with Tiny YOLO-V3, the detection speed is improved by 4 times, and the accuracy is improved by 2 percentage points.

Pothole Detection Using Yolo V3: A Deep Learning Approach

Potholes on roads constitute a serious problem for citizens acting as pedestrians furthermore as vehicular drivers. Government bodies which carries with it engineers and workers are responsible to detect damages on roads. Manually assessing every single a part of the road is very time- consuming, requires lots of manpower and hence it cannot be done efficiently. the tactic to repair this issue by automating the detection. The study focuses on collecting and analyzing the dataset of potholes to coach a convolutional neural network. the thing detection system tiny YOLOv3 is employed for detecting the potholes. the look of a system is identified which may be used for developing a mobile application for detection and presenting a visualized view of the potholes.

Ship Target Detection and Recognition Method on Sea Surface Based on Multi-Level Hybrid Network

This paper proposes a method of ship detection and recognition based on a multi-level hybrid network, designing a noise reducing and smoothing image enhancement algorithm based on multi-level two-dimensional template filter and three-layer pyramid structure. This work constructs an adaptive segmentation detection and ultra-lightweight target classification network model combining global and local image gray statistics. With a combination of traditional image processing and deep learning methods, the demand for computing and storage resources is reduced greatly. This method can detect and recognize the ship targets near the sea-sky-level quickly and has been verified by real flight camera data, and the accuracy rate is more than 90%. In comparison to the Tiny YOLOV3 network, the accuracy rate is reduced by 5%, but the calculation efficiency is increased by 50 times, and the parameters are reduced by 550 times.

Scaling up face masks detection with YOLO on a novel dataset

Face mask detection is a challenging research problem of computer vision because of the small sized area of face mask. The unavailability of proper datasets makes this problem even harder to crack. To address this bottleneck, we propose a novel face masks detection dataset consisting of 52,635 images with more than 50,000 tight bounding boxes and annotations for four different class labels namely, with masks, without masks, masks incorrectly, and mask area, which makes it a novel contribution for variety of face masks classification and detection tasks. Further, this dataset is tested with eight variants of the YOLO algorithm to determine its effectiveness. On the proposed dataset, original YOLO v4 achieved a mAP value of 71.69 % which was highest among all the original YOLO variants, tiny YOLO v4 achieved a mAP value of 57.71 % which was highest among all tiny variants. To propose new face masks detection algorithms that can perform with high accuracy in a limited computational resources environment, we selected four tiny variants of the YOLO algorithm and proposed new architectures modifications in their feature extraction networks that increased the overall performance and specifically, improved mAP by 4.12 % for tiny YOLO v3 and 2.54 % for tiny YOLO v4.

Mask Wearing Detection Algorithm Based on Improved Tiny YOLOv3

The new coronavirus spreads widely through droplets, aerosols and other carriers. Wearing a mask can effectively reduce the probability of being infected by the virus. Therefore, it is necessary to monitor whether people wear masks in public to prevent the virus from spreading further. However, there is no mature general mask wearing detection algorithm. Based on tiny YOLOv3 algorithm, this paper realizes the detection of face with mask and face without mask, and proposes an improvement to the algorithm. First, the loss function of the bounding box regression is optimized, and the original loss function is optimized as the Generalized Intersection over Union (GIoU) loss. Second, the network structure is improved, the residual unit is introduced into the backbone to increase the depth of the network and the detection of two scales is expanded to three. Finally, the size of anchor boxes is clustered based on [Formula: see text]-means algorithm. The experimental results on the constructed dataset show that, compared with the tiny YOLOv3 algorithm, the algorithm proposed in this paper improves the detection accuracy while maintaining high-speed inference ability.

Ensemble Method of Deep Learning, Color Segmentation, and Image Transformation to Track, Localize, and Count Cotton Bolls Using a Moving Camera in Real-Time

HighlightsAn ensemble method using color segmentation, deep learning, and image transformation was developed.Experiments were conducted to compare the method with other state-of-the-art tracking algorithms.The optimized ensemble method to track bolls achieved 94.4% accuracy using weakly trained tiny YOLOv2 models.The method achieved 7.6 frames per second and outperformed five other tracking methods.Abstract. In robotic applications, good perception can be computationally costly and create undesirable latency before a control decision is initiated. Most of the methods available for object detection deep learning are either fast with low accuracy or slow with high accuracy. Fast and accurate methods are necessary to track and localize objects such as cotton bolls that may be visible or occluded by each other or not well illuminated. In this study, an ensemble of a deep learning method and other image processing techniques was used to detect cotton bolls in-field on defoliated plants. In each image, a trained deep learning method, the YOLOv2 model, was used to detect open cotton bolls, and color segmentation was applied to confirm if the bolls detected by the YOLOv2 model were actually white to avoid false positives. Boll tracking was performed by following the spatial movement of good features on the edges of the bolls using the Lucas-Kanade algorithm. An image transformation algorithm was applied to the next image in case the previously detected boll was lost to retrieve the information of the missing boll. Each tracked and localized boll was stored and counted to give the total number of bolls detected. In this study, detection accuracy was sacrificed for image processing speed by using the YOLOv2 model. Detection accuracy was improved by using an ensemble method that combined image color segmentation, optical flow, and image transformation. This method was compared to eight other open-source methods implemented in OpenCV. The ensemble method detected and counted bolls at a speed of 7.6 fps with an accuracy of 94.4% using the Jetson TX2 embedded system to process 1K resolution images, outperforming the other OpenCV methods in various measurements. Keywords: Boll counting, Cotton, Cotton harvesting, DarkFlow, Darknet, Deep learning, Machine vision, YOLOv2.