Original Network Model Research Articles

SMYOLO: Lightweight Pedestrian Target Detection Algorithm in Low-Altitude Scenarios

Due to the arbitrariness of the drone’s shooting angle of view and camera movement and the limited computing power of the drone platform, pedestrian detection in the drone scene poses a greater challenge. This paper proposes a new convolutional neural network structure, SMYOLO, which achieves the balance of accuracy and speed from three aspects: (1) By combining deep separable convolution and point convolution and replacing the activation function, the calculation amount and parameters of the original network are reduced; (2) by adding a batch normalization (BN) layer, SMYOLO accelerates the convergence and improves the generalization ability; and (3) through scale matching, reduces the feature loss of the original network. Compared with the original network model, SMYOLO reduces the accuracy of the model by only 4.36%, the model size is reduced by 76.90%, the inference speed is increased by 43.29%, and the detection target is accelerated by 33.33%, achieving minimization of the network model volume while ensuring the detection accuracy of the model.

Investigation Into Recognition Algorithm of Helmet Violation Based on YOLOv5-CBAM-DCN

Recognition of safety helmets wearing by construction workers is a common target detection topic in applications of deep learning-based image processing. This paper provides a study of an enhanced YOLOv5-based method, in which the challenges caused by complicated construction environment backgrounds, dense targets, and the irregular shape of safety helmets are addressed. In a trunk network, feature extraction is more based on the target shape by using the Deformable Convolution Net instead of the conventional convolution; in the Neck, a Convolutional Block Attention Module is introduced to weaken feature extraction of complex backgrounds by giving weights to enhance the characterization ability of target features; and the original network’s Generalized Intersection over Union Loss is replaced by Distance Intersection over Union Loss to overcome the problem of erroneous location when the population is dense. The dataset for the training network is created by mixing open-source datasets with autonomous collecting to evaluate the effectiveness of the algorithm. We observed that the improved model has a detection accuracy of 91.6%, up 2.3% over the original network model, and a detection speed of 29 frames per second, which is compliant with most security cameras’ capture frame rate.

Deep Convolutional Neural Network Optimization for Defect Detection in Fabric Inspection.

This research is aimed to detect defects on the surface of the fabric and deep learning model optimization. Since defect detection cannot effectively solve the fabric with complex background by image processing, this research uses deep learning to identify defects. However, the current network architecture mainly focuses on natural images rather than the defect detection. As a result, the network architecture used for defect detection has more redundant neurons, which reduces the inference speed. In order to solve the above problems, we propose network pruning with the Bayesian optimization algorithm to automatically tune the network pruning parameters, and then retrain the network after pruning. The training and detection process uses the above-mentioned pruning network to predict the defect feature map, and then uses the image processing flow proposed in this research for the final judgment during fabric defect detection. The proposed method is verified in the two self-made datasets and the two public datasets. In the part of the proposed network optimization results, the Intersection over Union (IoU) of four datasets are dropped by 1.26%, 1.13%, 1.21%, and 2.15% compared to the original network model, but the inference time is reduced to 20.84%, 40.52%, 23.02%, and 23.33% of the original network model using Geforce 2080 Ti. Furthermore, the inference time is also reduced to 17.56%, 37.03%, 19.67%, and 22.26% using the embedded system AGX Xavier. After the image processing part, the accuracy of the four datasets can reach 92.75%, 94.87%, 95.6%, and 81.82%, respectively. In this research, Yolov4 is also trained with fabric defects, and the results showed this model are not conducive to detecting long and narrow fabric defects.

Research on Real-Time Infrared Image Fault Detection of Substation High-Voltage Lead Connectors Based on Improved YOLOv3 Network

Currently, infrared fault diagnosis mainly relies on manual inspection and low detection efficiency. This paper proposes an improved YOLOv3 network for detecting the working state of substation high-voltage lead connectors. Firstly, dilated convolution is introduced into the YOLOv3 backbone network to process low-resolution element layers, so as to enhance the network’s extraction of image features, promote function propagation and reuse, and improve the network’s recognition performance of small targets. Then the fault detection model of the infrared image of the high voltage lead connector is created and the optimal infrared image test data set is obtained through multi-scale training. Finally, the performance of the improved network model is tested on the data set. The test results show that the improved YOLOv3 network model has an average detection accuracy of 84.26% for infrared image faults of high-voltage lead connectors, which is 4.58% higher than the original YOLOv3 network model. The improved YOLOv3 network model has an average detection time of 0.308 s for infrared image faults of high-voltage lead connectors, which can be used for real-time detection in substations.

Multi-Sensor Fusion Module for Perceptual Target Recognition for Intelligent Machine Learning Visual Feature Extraction

This paper investigates the multi-sensor fusion perceptual target recognition based on machine learning visual feature extraction. On the basis of the efficient neural network model (Enet), an improved efficient semantic segmentation network model (Enet-CRF) is designed according to the requirements of the research content. This network model merges a CRF-RNN back-end optimization network to the original Enet, and further improves the classification performance of the network by adding constraints on the position relationship between image pixels and RGB information. The experiments indicate that the designed Enet-CRF has a certain improvement in obstacle classification performance, especially for small-scale obstacles such as pedestrians and bicycles. By deep fusion of LiDAR and vision sensors at the data level and higher-order features, this method adds high-precision radar information to the original Enet-CRF network model, thereby further enhancing the accuracy of obstacle classification and achieving an obstacle recognition method. This approach can not only accurately classify obstacles, but also attain accurate spatial information of obstacles, with good real-time performance.

Using ontologies to enhance human understandability of global post-hoc explanations of black-box models

The interest in explainable artificial intelligence has grown strongly in recent years because of the need to convey safety and trust in the ‘how’ and ‘why’ of automated decision-making to users. While a plethora of approaches has been developed, only a few focus on how to use domain knowledge and how this influences the understanding of explanations by users. In this paper, we show that by using ontologies we can improve the human understandability of global post-hoc explanations, presented in the form of decision trees. In particular, we introduce Trepan Reloaded, which builds on Trepan, an algorithm that extracts surrogate decision trees from black-box models. Trepan Reloaded includes ontologies, that model domain knowledge, in the process of extracting explanations to improve their understandability. We tested the understandability of the extracted explanations by humans in a user study with four different tasks. We evaluate the results in terms of response times and correctness, subjective ease of understanding and confidence, and similarity of free text responses. The results show that decision trees generated with Trepan Reloaded, taking into account domain knowledge, are significantly more understandable throughout than those generated by standard Trepan. The enhanced understandability of post-hoc explanations is achieved with little compromise on the accuracy with which the surrogate decision trees replicate the behaviour of the original neural network models.

ExYOLO: A small object detector based on YOLOv3 Object Detector

Object detection is a hot and popular research area in computer vision. Although object detection algorithm will be greatly speeding up with Deep Neural Networks, like YOLOv3 Object Detector, its performance in accuracy and precision was not satisfactory in small object detection which has less than 1% area in one image. This paper proposes a small object detector, exYOLO, which enhances the shallow feature information by superimposing the 2-fold down-sampled feature map from the original network model. The author attempted the set of experiments with a view to demonstrating accuracy and precision in small object detection, the results of the experiment indicate that exYOLO is a better performance in accuracy and precision.

Vehicle detection algorithm based on LW-SSD

Due to the complex structure of the traditional SSD network model and the large amount of parameters, it is not suitable for porting to mobile or embedded platforms, so the LW-SSD vehicle detection algorithm is proposed, this paper redesigns the basic feature extraction network of SSD, using MobileNet feature extraction network instead of VGG16, for the original MobileNet to improve, remove the original MobileNet network average pooling layer and Softmax layer, while reducing the MobileNet network from the original 28 layers to 13 layers. Then, on the basis of the subsequent network, feature maps of different scales are scaled, residual mapping and deep feature fusion operations, and the low-level position information and high-level semantic information are effectively combined to obtain new feature maps of different scales. The designed network model is trained using the Tensorflow framework, and the test effect of the network model is evaluated. The experimental results show that the accuracy of the network model designed in this paper is 78.76%, which is 9.27% lower than the original SSD network model. However, the amount of parameters is only 1/4 of the SSD network model. At the expense of partial accuracy, the design of a lightweight network model is exchanged, which lays the foundation for subsequent development on mobile or embedded platforms.

Structure Design of Convolutional Neural Network Based on Residual Theory for Face Recognition

Compared with the traditional face recognition methods, the deep convolution neural network model does not need to manually design complex and time-consuming feature extraction algorithms, but only needs to design an effective neural network model, and then carry out end-to-end, simple and efficient training on a large number of training samples, so as to obtain better classification accuracy. In this paper, based on the original VGG network model, combined with Residual theory, a deep-level residual convolutional neural network structure is designed and implemented, which not only reduces the computing power requirements of hardware computers, but also achieves better recognition results.

Producing Stable Periodic Solutions of Switched Impulsive Delayed Neural Networks Using a Matrix-Based Cubic Convex Combination Approach.

This article is dedicated to designing a novel periodic impulsive control strategy for producing globally exponentially stable periodic solutions for switched neural networks with discrete and finite distributed time-varying delays. First, tunable parameters and cubic convex combination approach are proposed to study the globally exponential convergence of switched neural networks. Second, a sufficient criterion for the existence, uniqueness, and globally exponential stability of a periodic solution is demonstrated by using contraction mapping theorem and the impulse-delay-dependent Lyapunov-Krasovskii functional method. It is worth emphasizing that the addressed Lyapunov-Krasovskii functional covers both triple integral terms and novel quadruple integral terms, which makes the conservatism of the above criteria decrease. Even if the original neural network models are unstable or the impulsive effects are strong, the addressed neural network model can produce a globally exponentially stable periodic solution. These results here, which include boundedness, globally uniformly exponential convergence, and globally exponentially stability of the periodic solution, generalize and improve the earlier publications. Finally, two numerical examples and their computer simulations are given to show the effectiveness of theoretical results.

INVESTIGATIONS ON REPRESENTATIVE ELEMENTARY VOLUME AND DIRECTIONAL PERMEABILITY OF FRACTAL-BASED FRACTURE NETWORKS USING POLYGON SUB-MODELS

Since the directional permeability of fractured rock masses is significantly dependent on the geometric properties of fractures, in this work, a numerical study was performed to analyze the relationships between them, in which fracture length follows a fractal distribution. A method to estimate the representative elementary volume (REV) size and directional permeability ([Formula: see text] by extracting regular polygon sub-models with different orientation angles ([Formula: see text] and side lengths ([Formula: see text] from an original discrete fracture network (DFN) model was developed. The results show that the fracture number has a power law relationship with the fracture length and the evolution of the exponent agrees well with that reported in previous studies, which confirms the reliability of the proposed fractal length distribution and stochastically generated DFN models. The [Formula: see text] varies significantly due to the influence of random numbers utilized to generate fracture location, orientation and length when [Formula: see text] is small. When [Formula: see text] exceeds some certain values, [Formula: see text] holds a constant value despite of [Formula: see text], in which the model scale is regarded as the REV size and the corresponding area of DFN model is represented by [Formula: see text] (in 2D). The directional permeability contours for DFN models plotted in the polar coordinate system approximate to circles when the model size is greater than the REV size. The [Formula: see text] decreases with the increment of fractal dimension of fracture length distribution ([Formula: see text]. However, the decreasing rate of [Formula: see text] (79.5%) when [Formula: see text] increases from 1.4 to 1.5 changes more significantly than that (34.8%) when [Formula: see text] increases from 1.5 to 1.6 for regular hexagon sub-models. This indicates that the small non-persistent fractures dominate the preferential flow paths; thereafter, the flow rate distribution becomes more homogeneous when [Formula: see text] exceeds a certain value (i.e. 1.5). A larger [Formula: see text] results in a denser fracture network and a stronger conductivity.

Pedestrian Detection Algorithm for Intelligent Vehicles in Complex Scenarios.

Pedestrian detection is an important aspect of the development of intelligent vehicles. To address problems in which traditional pedestrian detection is susceptible to environmental factors and are unable to meet the requirements of accuracy in real time, this study proposes a pedestrian detection algorithm for intelligent vehicles in complex scenarios. YOLOv3 is one of the deep learning-based object detection algorithms with good performance at present. In this article, the basic principle of YOLOv3 is elaborated and analyzed firstly to determine its limitations in pedestrian detection. Then, on the basis of the original YOLOv3 network model, many improvements are made, including modifying grid cell size, adopting improved k-means clustering algorithm, improving multi-scale bounding box prediction based on receptive field, and using Soft-NMS algorithm. Finally, based on INRIA person and PASCAL VOC 2012 datasets, pedestrian detection experiments are conducted to test the performance of the algorithm in various complex scenarios. The experimental results show that the mean Average Precision (mAP) value reaches 90.42%, and the average processing time of each frame is 9.6 ms. Compared with other detection algorithms, the proposed algorithm exhibits accuracy and real-time performance together, good robustness and anti-interference ability in complex scenarios, strong generalization ability, high network stability, and detection accuracy and detection speed have been markedly improved. Such improvements are significant in protecting the road safety of pedestrians and reducing traffic accidents, and are conducive to ensuring the steady development of the technological level of intelligent vehicle driving assistance.

Semantic combined network for zero‐shot scene parsing

Recently, image-based scene parsing has attracted increasing attention due to its wide application. However, conventional models can only be valid on images with the same domain of the training set and are typically trained using discrete and meaningless labels. Inspired by the traditional zero-shot learning methods which employ auxiliary side information to bridge the source and target domains, the authors propose a novel framework called semantic combined network (SCN), which aims at learning a scene parsing model only from the images of the seen classes while targeting on the unseen ones. In addition, with the assistance of semantic embeddings of classes, the proposed SCN can further improve the performances of traditional fully supervised scene parsing methods. Extensive experiments are conducted on the data set Cityscapes, and the results show that the proposed SCN can perform well on both zero-shot scene parsing (ZSSP) and generalised ZSSP settings based on several state-of-the-art scenes parsing architectures. Furthermore, the authors test the proposed model under the traditional fully supervised setting and the results show that the proposed SCN can also significantly improve the performances of the original network models.

Developing a Scene-Based Triangulated Irregular Network (TIN) Technique for Individual Tree Crown Reconstruction with LiDAR Data

LiDAR (Light Detection and Ranging)-based individual tree crown reconstruction is a challenge task due to the variable canopy morphologies and the penetrating properties of LiDAR to tree crown surfaces. Traditional methods, including LiDAR-derived rasterization, low-pass filtering smooth algorithm, and original triangular irregular network (TIN) model, have difficulties in balancing morphological accuracy and model smoothness. To address this issue, a scene-based TIN was generated with three steps based on the local scene principle. First, local Delaunay triangles were formed through connecting neighboring point sets. Second, key control points within each local Delaunay triangle, including steeple, inverted tip, ridge, saddle, and horseshoe shape control points, were extracted by analyzing multiple local scenes. These key points were derived to determine the fluctuations of forest canopies. Third, the scene-based TIN model was generated using the control points as nodes. Visual analysis indicates the new model can accurately reconstruct different canopy shapes with a relatively smooth surface, and statistical analysis of individual trees confirms that the overall error of the new model is smaller than others. Especially, the scene-based TIN derived raster reduced the average error to 0.18 m, with a standard deviation of 0.41, while the average errors of LiDAR-derived raster, low-pass filtered smooth raster, and original TIN derived raster have average errors of 0.96, 2.05, and 1.00 m, respectively. The local scene-based control point extraction also reduces data storage due to the elimination of redundant points, and furthermore the different point densities on different objects are beneficial for canopy segmentation.

Network model of passenger railway rolling stock scheduling

Passenger rolling stock scheduling to the passenger train timetable is based on the same type and number of carriages, which is the standard way of passenger trainset assignment. The possibility of train scheduling with a single carriage, a group of carriages or a railcar is topical issue for each carrier engaged in rail passenger transport in order to reduce the required rolling stock fleet. The paper presents an original network model developed for rolling stock scheduling for trainsets. The model makes possible to determine the optimum distribution of rolling stock in trainsets minimizing downtime at stations through the out-of-kilter algorithm. The model developed is applied to the scheduling of trainsets in part of the railway network in Bulgaria served by BDZ Passenger Transport Enterprise at the Plovdiv region. The use of the model allows to save passenger rolling stock and increases the efficiency of passenger rolling stock operation.

Automatic Melody Composition Using Enhanced GAN

In traditional music composition, the composer has a special knowledge of music and combines emotion and creative experience to create music. As computer technology has evolved, various music-related technologies have been developed. To create new music, a considerable amount of time is required. Therefore, a system is required that can automatically compose music from input music. This study proposes a novel melody composition method that enhanced the original generative adversarial network (GAN) model based on individual bars. Two discriminators were used to form the enhanced GAN model: one was a long short-term memory (LSTM) model that was used to ensure correlation between the bars, and the other was a convolutional neural network (CNN) model that was used to ensure rationality of the bar structure. Experiments were conducted using bar encoding and the enhanced GAN model to compose a new melody and evaluate the quality of the composition melody. In the evaluation method, the TFIDF algorithm was also used to calculate the structural differences between four types of musical instrument digital interface (MIDI) file (i.e., randomly composed melody, melody composed by the original GAN, melody composed by the proposed method, and the real melody). Using the TFIDF algorithm, the structures of the melody composed were compared by the proposed method with the real melody and the structure of the traditional melody was compared with the structure of the real melody. The experimental results showed that the melody composed by the proposed method had more similarity with real melody structure with a difference of only 8% than that of the traditional melody structure.

Influence of the topology on the dynamics of a complex network of HIV/AIDS epidemic models

In this paper, we propose an original complex network model for an epidemic problem in an heterogeneous geographical area. The complex network is constructed by coupling nonidentical instances of a HIV/AIDS epidemiological model for which a disease-free equilibrium and an endemic equilibrium can coexist. After proving the existence of a positively invariant region for the solutions of the complex network problem, we investigate the effect of the coupling on the dynamics of the network, and establish the existence of a unique disease-free equilibrium for the whole network, which is globally asymptotically stable. We prove the existence of an optimal topology that minimizes the level of infected individuals, and apply the theoretical results to the case of the Cape Verde archipelago.

A Bi-LSTM memory network for end-to-end goal-oriented dialog learning

We develop a model to satisfy the requirements of Dialog System Technology Challenge 6 (DSTC6) Track 1: building an end-to-end dialog systems for goal-oriented applications. This task involves learning a dialog policy from transactional dialogs in a given domain. Automatic system responses are generated using given task-oriented dialog data (http://workshop.colips.org/dstc6/index.html). As this task has a similar structure to a question answering task (Weston et al., 2015), we employ the MemN2N architecture (Sukhbaatar et al., 2015), which outperforms models based on recurrent neural networks or long short-term memory (LSTM). However, two problems arise when applying this model to the DSTC6 task. First, we encounter an out-of-vocabulary problem, which we resolve by categorizing the metadata types of words that exist in the knowledge base; the metadata is similar to the named entity. Second, the original memory network model has a weak ability to reflect sufficient temporal information, because it only uses sentence-level embeddings. Therefore, we add bidirectional LSTM (Bi-LSTM) at the beginning of the model to better reflect temporal information. The experimental results demonstrate that our model reflects temporal features well. Furthermore, our model achieves state-of-the-art performance among the memory networks, and is comparable to hybrid code networks (Ham et al., 2017) and hierarchical LSTM model (Bai et al., 2017) which is not an end-to-end architecture.

The effect of a loss of model structural detail due to network skeletonization on contamination warning system design: case studies

Abstract. The effect of limitations in the structural detail available in a network model on contamination warning system (CWS) design was examined in case studies using the original and skeletonized network models for two water distribution systems (WDSs). The skeletonized models were used as proxies for incomplete network models. CWS designs were developed by optimizing sensor placements for worst-case and mean-case contamination events. Designs developed using the skeletonized network models were transplanted into the original network model for evaluation. CWS performance was defined as the number of people who ingest more than some quantity of a contaminant in tap water before the CWS detects the presence of contamination. Lack of structural detail in a network model can result in CWS designs that (1) provide considerably less protection against worst-case contamination events than that obtained when a more complete network model is available and (2) yield substantial underestimates of the consequences associated with a contamination event. Nevertheless, CWSs developed using skeletonized network models can provide useful reductions in consequences for contaminants whose effects are not localized near the injection location. Mean-case designs can yield worst-case performances similar to those for worst-case designs when there is uncertainty in the network model. Improvements in network models for WDSs have the potential to yield significant improvements in CWS designs as well as more realistic evaluations of those designs. Although such improvements would be expected to yield improved CWS performance, the expected improvements in CWS performance have not been quantified previously. The results presented here should be useful to those responsible for the design or implementation of CWSs, particularly managers and engineers in water utilities, and encourage the development of improved network models.

The effect of a loss of model structural detail due to network skeletonization on contamination warning system design: case studies.

The effect of limitations in the structural detail available in a network model on contamination warning system (CWS) design was examined in case studies using the original and skeletonized network models for two water distribution systems (WDSs). The skeletonized models were used as proxies for incomplete network models. CWS designs were developed by optimizing sensor placements for worst-case and mean-case contamination events. Designs developed using the skeletonized network models were transplanted into the original network model for evaluation. CWS performance was defined as the number of people who ingest more than some quantity of a contaminant in tap water before the CWS detects the presence of contamination. Lack of structural detail in a network model can result in CWS designs that (1) provide considerably less protection against worst-case contamination events than that obtained when a more complete network model is available and (2) yield substantial underestimates of the consequences associated with a contamination event. Nevertheless, CWSs developed using skeletonized network models can provide useful reductions in consequences for contaminants whose effects are not localized near the injection location. Mean-case designs can yield worst-case performances similar to those for worst-case designs when there is uncertainty in the network model. Improvements in network models for WDSs have the potential to yield significant improvements in CWS designs as well as more realistic evaluations of those designs. Although such improvements would be expected to yield improved CWS performance, the expected improvements in CWS performance have not been quantified previously. The results presented here should be useful to those responsible for the design or implementation of CWSs, particularly managers and engineers in water utilities, and encourage the development of improved network models.