YOLOv3 Model Research Articles

Multi-Target Defect Identification for Railway Track Line Based on Image Processing and Improved YOLOv3 Model

The condition monitoring of railway track line is one of the essential tasks to ensure the safety of the railway transportation system. Railway track line is mainly composed of tracks, fasteners, sleepers, and so on. Given the requirements for rapid and accurate inspection, innovative and intelligent methods for multi-target defect identification of the railway track line using image processing and deep learning methods are proposed in this paper. Firstly, the track and fastener positioning method based on variance projection and wavelet transform is introduced. After that, a bag-of-visual-word (BOVW) model combined with spatial pyramid decomposition is proposed for railway track line multi-target defect detection with a detection accuracy of 96.26%. Secondly, an improved YOLOv3 model named TLMDDNet (Track Line Multi-target Defect Detection Network), integrating scale reduction and feature concatenation, is proposed to enhance detection accuracy and efficiency. Finally, to reduce model complexity and further improve the detection speed, with the help of dense connection structure, a lightweight design strategy for the TLMDDNet model named DC-TLMDDNet (Dense Connection Based TLMDDNet) is proposed, in which the DenseNet is applied to optimize feature extraction layers in the backbone network of TLMDDNet. The effectiveness of the proposed methods is demonstrated by the experimental results.

Effects of Gradient Optimizer on Model Pruning

Deep convolutional neural networks deploy in applications are hindered by their large parameters and high computational cost. In this paper, we test different gradient optimizer’s effect on YOLOv3 model pruning. This is achieved by enforcing channel-level sparsity in YOLOv3 network. The model trained with Adam optimizer get 5× reduction in model size only after training 60 epochs.

基于改进YOLOv2模型的多目标识别方法

基于改进YOLOv2模型的多目标识别方法

Detection of Collapsed Buildings in Post-Earthquake Remote Sensing Images Based on the Improved YOLOv3

An important and effective method for the preliminary mitigation and relief of an earthquake is the rapid estimation of building damage via high spatial resolution remote sensing technology. Traditional object detection methods only use artificially designed shallow features on post-earthquake remote sensing images, which are uncertain and complex background environment and time-consuming feature selection. The satisfactory results from them are often difficult. Therefore, this study aims to apply the object detection method You Only Look Once (YOLOv3) based on the convolutional neural network (CNN) to locate collapsed buildings from post-earthquake remote sensing images. Moreover, YOLOv3 was improved to obtain more effective detection results. First, we replaced the Darknet53 CNN in YOLOv3 with the lightweight CNN ShuffleNet v2. Second, the prediction box center point, XY loss, and prediction box width and height, WH loss, in the loss function was replaced with the generalized intersection over union (GIoU) loss. Experiments performed using the improved YOLOv3 model, with high spatial resolution aerial remote sensing images at resolutions of 0.5 m after the Yushu and Wenchuan earthquakes, show a significant reduction in the number of parameters, detection speed of up to 29.23 f/s, and target precision of 90.89%. Compared with the general YOLOv3, the detection speed improved by 5.21 f/s and its precision improved by 5.24%. Moreover, the improved model had stronger noise immunity capabilities, which indicates a significant improvement in the model’s generalization. Therefore, this improved YOLOv3 model is effective for the detection of collapsed buildings in post-earthquake high-resolution remote sensing images.

Study on Holographic Image Recognition Technology of Zooplankton

Marine zooplankton has important ecological and economic value. The observation and automatic image recognition technology of marine zooplankton is an important mean to acquire data such as species, quantity, spatial distribution and behavioral postures of zooplankton, and is an important support for marine scientific research. Digital holography has an innate advantage of refocusing and reconstruction, which is suitable for deep learning and living zooplankton recognition. In this study, a large number of holographic images was trained by using the improved YOLOv2 model, and after test, the study achieved satisfactory results: the models trained by the images with sharpness assessment score of 0.6 or higher, have precision rate above 94% and a recall rate above 88%. This study mainly discusses: (1) the detection method of moving targets to acquire the images of moving zooplankton; (2) the two factors that affect the holographic images recognition results, mean (pixel mean of images) subtraction operation and image sharpness, and the no-reference sharpness assessment based on structural similarity for holographic images; (3) the relationship between sharpness assessment index or mean subtraction and the recognition results.

The Research of Gait Recognition Based on Deep Learning: A Case Study of the Missing Elderly

In our daily life, we often see some notices for the missing elderly. As we all know, the elderly people with Alzheimer’s disease or memory impairment are more likely and more often to get lost. Therefore, I hope that I can make some contribution to help solve this problem. For this reason, the author a detection and tracking system with interactive interface was designed by the author based on gait recognition for the elder people who are under the risk of getting lost. The walking characteristics of the elderly can be extracted from the gait image data of the elderly provided by their families. Combined with the video streaming information documented by the cameras of the public security organ, the detection and tracking system can identify the missing elderly at a long distance, in this way, it provides clues for searching the missing elderly. In this paper, all experiments were under the Windows 7 operating system. Python was used to invocate yolov3 model for character detection. The foreground images were obtained by applying background subtraction in the static background, and then gray processing, Gauss blurring, binarization and other image operations were done to extract the image features and obtain the GEI gait energy map with the help of gait feature extraction technology. The image data were stored in the database, and then pedestrian identity detection was carried out through an input video or by real-time cameras. In order to test conveniently, an interactive interface based on PyQt4 is designed, which can realize functions such as register pedestrian names on site, acquire pedestrian gait energy map and establish database, refresh background, detect and recognize pedestrians. Moreover, it supports arbitrary switching between real-time camera detection and input video detection.

Real-time behavior detection and judgment of egg breeders based on YOLO v3

In order to detect the behavior of egg breeders in self-breeding cages rapidly, a method of target location and behavior recognition based on visual images was proposed. In this study, Hy-Line Gray chickens were bred as objects. Through manual marking, the training set, validation set and test set were established, and YOLO v3 model was adopted to detect the collected images. The value of subdivision and batch size were determined by experiment. The learning rate was dynamically adjusted according to the change of loss value in the training model. Finally, the mean average precision of the trained model on the validation set was 92.09%. In this paper, the recognition rates of six kinds of behaviors in the morning and in the afternoon and under different densities were analyzed. Furthermore, a kind of welfare indicator was tested and abnormal behavior was evaluated. The results showed that: The mean precision rate of the six behaviors was followed by mating (94.72%), stand (94.57%), feed (93.10%), spread (92.02%), fight (88.67%) and drink (86.88%). The mean false rate ranged from low to high was spread (0.11%), mating (0.14%), fight (0.20%), drink (0.25%), feed (1.17%) and stand (8.62%). The mean missing rate ranged from low to high was mating (4.65%), stand (5.01%), feed (5.15%), spread (6.25%), fight (14.69%) and drink (15.79%). The method presented in this paper has a good effect on identifying the behavior of egg breeders, which can provide technical support for the promotion of the self-breeding mode.

Deep Learning for Aircraft Wake Vortex Identification

Aircraft wake vortex refers to a pair of closed vortices due to the air pressure difference between the upper and lower surfaces of the wing when the aircraft is flying. During the take-off and landing of the aircraft, the wake vortex of the preceding aircraft threatened the safety of the aircraft following its approach. Especially in the near-field phase, due to the low altitude of the aircraft, the wake vortex is more harmful under the influence of the ground effect. Therefore, the wake vortex of the front aircraft can not be ignored. In order to maintain air traffic safety and accelerate the smooth flow of air traffic, this paper used a YOLO v3 model based on the principle of deep learning to identify the wake vortex generated by the aircraft. Through laser radar scanning, data processing, and deep neural network simulation experiments, the results show that the model can achieve high confidence recognition of the aircraft wake vortex, and can provide air traffic controllers with auxiliary decision information in actual work to ensure aircraft safety.

A YOLOv3-based non-helmet-use detection for seafarer safety aboard merchant ships

International shipping industry undertaking more than 80% global trade volume, is indispensable to the world’s economy and development. Seafarers, which are an important part of shipping, mainly engage in safe navigation and good maintenance of ships. However, they are prone to expose themselves to occupational accidents and injuries, especially fatal trauma on the head due to non-helmet-use behaviour. Wearing a safety helmet can effectively reduce the risks resulting in head injuries. Despite the vital role of safety helmet, many seafarers tend to take off their helmet because of discomfort caused by weight and high temperature in engine room or deck. To improve the ship’s safe operation and prevent the loss of life at sea, this paper proposed a state-of-the-art method based on YOLOv3 model for automatic and real-time detection of non-helmet-use behaviour. The experimental results showed that the detection method can run real-time under ship’s video surveillance and precisely detect the non-helmet-use behaviour with low miss rate.

Determination of Vehicle Trajectory through Optimization of Vehicle Bounding Boxes Using a Convolutional Neural Network.

In this manuscript, a new method for the determination of vehicle trajectories using an optimal bounding box for the vehicle is developed. The vehicle trajectory is extracted using images acquired from a camera installed at an intersection based on a convolutional neural network (CNN). First, real-time vehicle object detection is performed using the YOLOv2 model, which is one of the most representative object detection algorithms based on CNN. To overcome the inaccuracy of the vehicle location extracted by YOLOv2, the trajectory was calibrated using a vehicle tracking algorithm such as a Kalman filter and intersection-over-union (IOU) tracker. In particular, we attempted to correct the vehicle trajectory by extracting the center position based on the geometric characteristics of a moving vehicle according to the bounding box. The quantitative and qualitative evaluations indicate that the proposed algorithm can detect the trajectories of moving vehicles better than the conventional algorithm. Although the center points of the bounding boxes obtained using the existing conventional algorithm are often outside of the vehicle due to the geometric displacement of the camera, the proposed technique can minimize positional errors and extract the optimal bounding box to determine the vehicle location.

CORIGAN: Assessing multiple species and interactions within images

Abstract Images are resourceful data for ecologists and can provide a more complete information than other methods to study biodiversity and the interactions between species. Automated image analysis however often relies on extensive datasets, not implementable by small research teams. We are here proposing an object detection method that allows the analysis of high‐resolution images containing many animals interacting in a small dataset. We developed an image analysis pipeline named ‘CORIGAN’ to extract the characteristics of animal communities. CORIGAN is based on the YOLOv3 model as the core of object detection. To illustrate potential applications, we use images collected during a sentinel prey experiment. Our pipeline can be used to detect, count and study the physical interactions between various animals. On our example dataset, the model reaches 86.6% precision and 88.9% recall at the species level or even at the caste level for ants. The training set required fewer than 10 hr of labelling. Based on the pipeline output, it was possible to build the trophic and non‐trophic interactions network describing the studied community. CORIGAN relies on generic properties of the detected animals and can be used for a wide range of studies and supports. Here, we study invertebrates on high‐resolution images, but the same processing can be transferred for the study of larger animals on satellite or aircraft images.

Vehicle target detection in complex scenes based on YOLOv3 algorithm

In view of the low accuracy of traditional vehicle target detection methods in complex scenes, combined with the current hot development of deep learning, this paper applies the YOLOv3 algorithm framework to achieve vehicle target detection. By using PASCAL VOC2007 and VOC2012 data sets, images containing vehicle targets were screened out to constitute the VOC car data set, and the target detection problem was converted into a binary classification problem. Then loading the pre-trained YOLOv3 model weight, and training the vehicle target detection model weight based on YOLOv3 algorithm, which is used to detect the test samples. Experimental results show that this method has advantages over the traditional target detection algorithms in recognition accuracy and detection speed.

A novel YOLOv3-arch model for identifying cholelithiasis and classifying gallstones on CT images.

Locating diseases precisely from medical images, like ultrasonic and CT images, have been one of the most challenging problems in medical image analysis. In recent years, the vigorous development of deep learning models have greatly improved the accuracy in disease location on medical images. However, there are few artificial intelligent methods for identifying cholelithiasis and classifying gallstones on CT images, since no open source CT images dataset of cholelithiasis and gallstones is available for training the models and verifying their performance. In this paper, we build up the first medical image dataset of cholelithiasis by collecting 223846 CT images with gallstone of 1369 patients. With these CT images, a neural network is trained to “pick up” CT images of high quality as training set, and then a novel Yolo neural network, named Yolov3-arch neural network, is proposed to identify cholelithiasis and classify gallstones on CT images. Identification and classification accuracies are obtained by 10-fold cross-validations. It is obtained that our Yolov3-arch model is with average accuracy 92.7% in identifying granular gallstones and average accuracy 80.3% in identifying muddy gallstones. This achieves 3.5% and 8% improvements in identifying granular and muddy gallstones to general Yolo v3 model, respectively. Also, the average cholelithiasis identifying accuracy is improved to 86.50% from 80.75%. Meanwhile, our method can reduce the misdiagnosis rate of negative samples by the object detection model.

New approach to vehicle license plate location based on new model YOLO‐L and plate pre‐identification

Currently, the conventional license plate location method fails to detect the license plate under complex road environments such as severe weather conditions and viewpoint changes. Besides, it is difficult for license plate location method based on machine learning to precisely locate the area of license plate. Moreover, license plate location method may incorrectly detect similar objects such as billboards and road signs as license plates. To alleviate these problems, this article proposes a new approach to vehicle license plate location based on new model YOLO-L and plate pre-identification. The new model improves in two aspects to precisely locate the area of license plate. First, it uses k-means++ clustering algorithm to select the best number and size of plate candidate boxes. Second, it modifies the structure and depth of YOLOv2 model. Plate pre-identification algorithm can effectively distinguish license plates from similar objects. The experimental results show that authors' proposed method not only achieves a precision of 98.86% and a recall of 98.86%, which outperforms the existing methods, but also has high efficiency in real time.

Detection of Apple Lesions in Orchards Based on Deep Learning Methods of CycleGAN and YOLOV3-Dense

Plant disease is one of the primary causes of crop yield reduction. With the development of computer vision and deep learning technology, autonomous detection of plant surface lesion images collected by optical sensors has become an important research direction for timely crop disease diagnosis. In this paper, an anthracnose lesion detection method based on deep learning is proposed. Firstly, for the problem of insufficient image data caused by the random occurrence of apple diseases, in addition to traditional image augmentation techniques, Cycle-Consistent Adversarial Network (CycleGAN) deep learning model is used in this paper to accomplish data augmentation. These methods effectively enrich the diversity of training data and provide a solid foundation for training the detection model. In this paper, on the basis of image data augmentation, densely connected neural network (DenseNet) is utilized to optimize feature layers of the YOLO-V3 model which have lower resolution. DenseNet greatly improves the utilization of features in the neural network and enhances the detection result of the YOLO-V3 model. It is verified in experiments that the improved model exceeds Faster R-CNN with VGG16 NET, the original YOLO-V3 model, and other three state-of-the-art networks in detection performance, and it can realize real-time detection. The proposed method can be well applied to the detection of anthracnose lesions on apple surfaces in orchards.

Apple detection during different growth stages in orchards using the improved YOLO-V3 model

Real-time detection of apples in orchards is one of the most important methods for judging growth stages of apples and estimating yield. The size, colour, cluster density, and other growth characteristics of apples change as they grow. Traditional detection methods can only detect apples during a particular growth stage, but these methods cannot be adapted to different growth stages using the same model. We propose an improved YOLO-V3 model for detecting apples during different growth stages in orchards with fluctuating illumination, complex backgrounds, overlapping apples, and branches and leaves. Images of young apples, expanding apples, and ripe apples are initially collected. These images are subsequently augmented using rotation transformation, colour balance transformation, brightness transformation, and blur processing. The augmented images are used to create training sets. The DenseNet method is used to process feature layers with low resolution in the YOLO-V3 network. This effectively enhances feature propagation, promotes feature reuse, and improves network performance. After training the model, the performance of the trained model is tested on a test dataset. The test results show that the proposed YOLOV3-dense model is superior to the original YOLO-V3 model and the Faster R-CNN with VGG16 net model, which is the state-of-art fruit detection model. The average detection time of the model is 0.304s per frame at 3000 × 3000 resolution, which can provide real-time detection of apples in orchards. Moreover, the YOLOV3-dense model can effectively provide apple detection under overlapping apples and occlusion conditions, and can be applied in the actual environment of orchards.

Tender Tea Shoots Recognition and Positioning for Picking Robot Using Improved YOLO-V3 Model

To recognize the tender shoots for high-quality tea and to determine the picking points accurately and quickly, this paper proposes a method of recognizing the picking points of the tender tea shoots with the improved YOLO-v3 deep convolutional neural network algorithm. This method realizes the end-to-end target detection and the recognition of different postures of high-quality tea shoots, considering both efficiency and accuracy. At first, in order to predict the category and position of tender tea shoots, an image pyramid structure is used to obtain the characteristic map of tea shoots at different scales. The residual network block structure is added to the downsampling part, and the fully connected part is replaced by a $1\times 1$ convolution operation at the end, ensuring accurate identification of the result and simplifying the network structure. The K-means method is used to cluster the dimension of the target box. Finally, the image data set of picking points for high-quality tea shoots is built. The accuracy of the trained model under the verification set is over 90%, which is much higher than the detection accuracy of the research methods.

Fast vehicle logo detection in complex scenes

Vehicle logo detection has received considerable attention due to its many applications, such as brand reputation measurement and vehicle monitoring. Deep-learning algorithms, such as you only look once (YOLO) and faster-regional convolutional neural network, have obtained excellent performance for object detection. However, it is not easy to exploit these methods to detect small objects. This work presents a novel method for fast and accurate vehicle logo detection (VLD) in complex scenes. First, we modify the original YOLOv3 model for the VLD task and solve the small object detection problem by hard example training. Second, we construct a new VLD dataset known as VLD-30, which encourages us to develop a data-driven training method and improve the detection accuracy. Experimental results demonstrate that the proposed data-training method is useful and the modified YOLOv3 is effective for fast and accurate vehicle logo detection in complex scenes.

Research on Automatic Garbage Detection System Based on Deep Learning and Narrowband Internet of Things

This paper proposes an automatic garbage detection system based on deep learning and narrowband Internet of things. The system automatically detects and identifies decoration garbage directly in front-end embedded monitoring module, and manages thousands of monitoring front-ends through narrow-band Internet of Things and background server. In the front-end embedded module of the system, the improved YOLOv2 network model is adopted to do garbage detection and recognition. Means of target box dimension clustering and classification network pre-training are used to improve the YOLOv2 model performance; at the same time, the network is lightweight by replacing the feature extraction network and other methods, and the lightweight YOLOv2 network is optimized and ported to embedded module. The experiment test shows that compared with the traditional monitoring system, the cost is reduced by more than half, which can effectively save manpower and material resources, and the accuracy of detection and recognition has also been improved.

Recognition of Anthracnose Injuries on Apple Surfaces using YOLOV 3-Dense

Plant ailment is one of the essential drivers of harvest yield decrease. With the advancement of PC vision and profound learning innovation, independent discovery of plant surface sore pictures gathered by optical sensors has become a significant research bearing for convenient yield ailment analysis. Right now, anthracnose injury identification strategy dependent on profound learning is proposed. Right off the bat, for the issue of lacking picture information brought about by the irregular event of apple illnesses, notwithstanding conventional picture expansion strategies, Cycle-Consistent Adversarial Network (CycleGAN) profound learning model is utilized right now achieve information increase. These strategies adequately enhance the decent variety of preparing information and give a strong establishment to preparing the identification model. Right now, the premise of picture information increase, thickly associated neural system (DenseNet) is used to streamline highlight layers of the YOLO-V3 model which have lower goals. DenseNet extraordinarily improves the usage of highlights in the neural system and upgrades the identification consequence of the YOLO-V3 model. It is checked in tests that the improved model surpasses Faster R-CNN with VGG16 NET, the first YOLO-V3 model, and other three cutting edge arranges in discovery execution, and it can understand continuous recognition. The proposed technique can be all around applied to the recognition of anthracnose injuries on apple surfaces in-plantations.