Target Classification Accuracy Research Articles

Deep learning and quantization for accurate and efficient multi-target radar inference of moving targets

Real-time, radar-based human activity and target classification is useful for wide-area ground surveillance. However, the feasibility of deploying deep learning (DL) models in radar-based systems with limited computational resources remains unexplored. This paper investigated the effect of quantization on model throughput and accuracy for deployment in radar systems. A seven-layer residual network was proposed to classify ground-moving targets and achieved a test accuracy of 87.72%. The model was then quantized to 16-bit and 8-bit precision, resulting in a 3.8 times speedup in inference throughput, with less than a 0.4% drop in test and validation accuracy. The results showed that quantization can improve inference throughput with a negligible decrease in target classification accuracy. The increase in throughput and reduction in computational expense that comes with quantization promotes the feasibility of the deployment of DL models in systems with limited computational resources. The findings of this paper hold significant promise for the successful use of quantized models in modern radar systems, while adhering to stringent size, weight and power consumption constraints.

Classification and Recognition Method of Non-Cooperative Objects Based on Deep Learning.

Accurately classifying and identifying non-cooperative targets is paramount for modern space missions. This paper proposes an efficient method for classifying and recognizing non-cooperative targets using deep learning, based on the principles of the micro-Doppler effect and laser coherence detection. The theoretical simulations and experimental verification demonstrate that the accuracy of target classification for different targets can reach 100% after just one round of training. Furthermore, after 10 rounds of training, the accuracy of target recognition for different attitude angles can stabilize at 100%.

A Neural Network-Based Weighted Voting Algorithm for Multi-Target Classification in WSN.

One of the most important applications in the wireless sensor networks (WSN) is to classify mobile targets in the monitoring area. In this paper, a neural network(NN)-based weighted voting classification algorithm is proposed on the basis of the NN-based classifier and combined with the idea of voting strategy, which is implemented on the nodes of the WSN monitoring system by means of the "upper training, lower transplantation" approach. The performance of the algorithm is verified by using real-world experimental data, and the results show that the proposed method has a higher accuracy in classifying the target signal features, achieving an average classification accuracy of about 85% when utilizing a deep neural network (DNN) and deep belief network (DBN) as the base classifier. The experiment reveals that the NN-based weighted voting algorithm enhances the target classification accuracy by approximately 5% in comparison to the single NN-based classifier, but the memory and computation time required for the algorithm to run are also increased at the same time. Compared to the FFNN classifier, which exhibited the highest classification accuracy among the four selected methods, the algorithm achieves an improvement of approximately 8.8% in classification accuracy. However, it incurs greater overhead time to run.

Simulation-based training of neural networks for low-frequency sonar target classification

Artificial Intelligence (AI) is widely used in various fields, but obtaining sufficient underwater acoustic data for AI training remains challenging. Even if simulation data are used for training, the classification accuracy could be lowered due to changes in the environment around the target. This study presents the research findings on training neural networks for target classification using data generated through simulations. We trained neural networks using target scattering signals obtained through finite element analysis for various environments. Subsequently, we utilized trained neural networks to classify the spherical shells on the seabed acquired in water tank experiments. To solve the problem of lack of data for AI training, we generated training data by applying realistic variations of parameters such as sound speed of water, target material properties, and thickness. Such training data provide more meaningful information than noise-added data. The targets considered were spherical shells with and without internal spaces. Although the target classification accuracy is not yet high enough, it is expected to be used to overcome the lack of data by applying to untrained targets. [Work supported by KIMST funded by the Agency of Korea Coast Guard (KIMST-20210547) and by KRISO funded by Ministry of Oceans and Fisheries (PES4380).]

1D CNN-GRU 기반 불특정 비행 거리 궤적을 가진 미사일 분류를 위한 학습 방법에 관한 연구

In this paper, we propose a training method and an optimal training distance interval to improve target classification performance based on the one-dimensional convolutional neural network-gated recurrent unit (1D CNN-GRU). To improve classification performance, we describe two training methods and confirm classification performance. The structure and setting parameters of 1D CNN-GRU use the results optimized in the previous study. We analyze classification performance at intervals of 10km from 20 km to 80 km by training the network. As a result of applying the proposed training method, the high classification performance at all training distances for unfixed flight range trajectories was obtained at over 89.53 % in the absence of noise and over 80.38 % in the case of severe noise, respectively. In addition, considering the training efficiency and target classification accuracy, it was confirmed that 50 km was suitable for training distance intervals with the average classification accuracy of 99.70 % in non-noise and 89.19 % in the case of severe noise.

Research on Metal Target Classification Method Based on Multi-Frequency Electromagnetic Induction

The traditional single-frequency metal mine detector is limited by the single-frequency component, which is difficult to meet the requirements of high-accuracy detection of multiple types of mine targets. The modulation square wave is used to transmit the multi-frequency signal based on the multi-frequency electromagnetic induction technology. The phase-locked amplifier is used to extract the signal feature of each frequency component. The naive bayesian algorithm is used to classify metal targets according to the extracted feature vectors. This method can effectively improve target classification accuracy compared with single-frequency metal mine detectors.

Target Recognition in SAR Images by Deep Learning with Training Data Augmentation.

Mass production of high-quality synthetic SAR training imagery is essential for boosting the performance of deep-learning (DL)-based SAR automatic target recognition (ATR) algorithms in an open-world environment. To address this problem, we exploit both the widely used Moving and Stationary Target Acquisition and Recognition (MSTAR) SAR dataset and the Synthetic and Measured Paired Labeled Experiment (SAMPLE) dataset, which consists of selected samples from the MSTAR dataset and their computer-generated synthetic counterparts. A series of data augmentation experiments are carried out. First, the sparsity of the scattering centers of the targets is exploited for new target pose synthesis. Additionally, training data with various clutter backgrounds are synthesized via clutter transfer, so that the neural networks are better prepared to cope with background changes in the test samples. To effectively augment the synthetic SAR imagery in the SAMPLE dataset, a novel contrast-based data augmentation technique is proposed. To improve the robustness of neural networks against out-of-distribution (OOD) samples, the SAR images of ground military vehicles collected by the self-developed MiniSAR system are used as the training data for the adversarial outlier exposure procedure. Simulation results show that the proposed data augmentation methods are effective in improving both the target classification accuracy and the OOD detection performance. The purpose of this work is to establish the foundation for large-scale, open-field implementation of DL-based SAR-ATR systems, which is not only of great value in the sense of theoretical research, but is also potentially meaningful in the aspect of military application.

G2Grad-CAMRL: An Object Detection and Interpretation Model Based on Gradient-Weighted Class Activation Mapping and Reinforcement Learning in Remote Sensing Images

Remote sensing images contain important information such as airports, ports, and ships. By extracting remote sensing image features and learning the mapping relationship between image features and text semantic features, the interpretation and description of remote sensing image content can be realized, which has a wide range of application value in military and civil fields such as national defense security, land monitoring, urban planning, disaster mitigation and so on. Aiming at the complex background of remote sensing images and the lack of interpretability of existing target detection models, and the problems in feature extraction between different network structures, different layers, and the accuracy of target classification, we propose an object detection and interpretation model based on Gradient weighted class activation mapping and reinforcement learning. Firstly, ResNet is used as the main backbone network to extract the features of remote sensing images and generate feature graphs. Then, we add the global average pooling layer to obtain the corresponding feature weight vector of the feature graph. The weighted vectors are superimposed to output class activation maps. The reinforcement learning method is used to optimize the generated region generation network. At the same time, we improve the reward function of reinforcement learning to improve the effect of the region generation network. Finally, network dissecting analysis is used to obtain the interpretable semantic concept in the model. Through experiments, the average accuracy is more than 85%. Experimental results in the public remote sensing image description dataset show that the proposed method has high detection accuracy and good description performance for remote sensing images in complex environments.

Underwater Image Classification Algorithm Based on Convolutional Neural Network and Optimized Extreme Learning Machine

In order to deal with the target recognition in the complex underwater environment, we carried out experimental research. This includes filtering noise in the feature extraction stage of underwater images rich in noise, or with complex backgrounds, and improving the accuracy of target classification in the recognition process. This paper discusses our contribution to improving the accuracy of underwater target classification. This paper proposes an underwater target classification algorithm based on the improved flow direction algorithm (FDA) and search agent strategy, which can simultaneously optimize the weight parameters, bias parameters, and super parameters of the extreme learning machine (ELM). As a new underwater target classifier, it replaces the full connection layer in the traditional classification network to build a classification network. In the first stage of the network, the DenseNet201 network pre-trained by ImageNet is used to extract features and reduce dimensions of underwater images. In the second stage, the optimized ELM classifier is trained and predicted. In order to weaken the uncertainty caused by the random input weight and offset of the introduced ELM, the fuzzy logic, chaos initialization, and multi population strategy-based flow direction algorithm (FCMFDA) is used to adjust the input weight and offset of the ELM and optimize the super parameters with the search agent strategy at the same time. We tested and verified the FCMFDA-ELM classifier on Fish4Knowledge and underwater robot professional competition 2018 (URPC 2018) datasets, and achieved 99.4% and 97.5% accuracy, respectively. The experimental analysis shows that the FCMFDA-ELM underwater image classifier proposed in this paper has a greater improvement in classification accuracy, stronger stability, and faster convergence. Finally, it can be embedded in the recognition process of underwater targets to improve the recognition performance and efficiency.

Sunken oil detection and classification using MBES backscatter data

Sunken oil incidents have occurred multiple times in the Bohai Sea over the past ten years. Currently, quick and effective sunken oil detection and classification remains a difficult problem. In this study, sonar detection experiments are conducted to obtain acoustic image samples using a multibeam echosounder (MBES) in a large seawater tank at the bottom of the area where the sunken oil is located. A series of MBES data corrections are constructed to generate backscatter strength images that can reflect the target characteristics directly. Meanwhile, eight-dimensional features are extracted, and a support vector machine (SVM) classification framework is built to classify the sunken oil and other interference targets. The results indicate that the MBES backscatter images provide an alternative approach for detecting and classifying sunken oil. The overall target classification accuracy reaches 88.5% by the SVM algorithm. Thus, this study provides a basis for further investigation of detecting sunken oil.

A Semantic Image Retrieval Method Based on Interest Selection

There is a semantic gap between people’s understanding of images and the underlying visual features of images, which makes it difficult for image retrieval results to meet the needs of individual interests. To overcome the semantic gap in image retrieval, this paper proposes a semantic image retrieval method based on interest selection. This method analyses the interest points of individual selections and gives the weight of the interest selection in different regions of an image. By extracting the underlying visual features of different regions, this paper constructs two feature vector methods after users’ interest point weighting. The two methods are called interest weighted summation and interest weighting. Finally, this paper compares the accuracy of different image classification methods using a support vector machine classification algorithm. The experimental results show that the target classification accuracy of the classification algorithm based on interest weighted summation is higher than that of the traditional and interest weighted methods. The classification algorithm based on interest weighted summation has the highest overall effect on target object classification in the four experimental scenarios. Therefore, the interest point selection method can effectively improve the overall satisfaction of image recommendation and can be used as a novel solution to overcome the semantic gap.

Optimal Data Collection Period for Continuous Glucose Monitoring to Assess Long-Term Glycemic Control: Revisited.

It is not clear how the short-term continuous glucose monitoring (CGM) sampling time could influence the bias in estimating long-term glycemic control. A large bias could, in the worst case, lead to incorrect classification of patients achieving glycemic targets, nonoptimal treatment, and false conclusions about the effect of new treatments. This study sought to investigate the relation between sampling time and bias in the estimates. We included a total of 329 type 1 patients (age 14-86 years) with long-term CGM (90 days) data from three studies. The analysis calculated the bias from estimating long-term glycemic control based on short-term sampling. Time in range (TIR), time above range (TAR), time below range (TBR), correlation, and glycemic target classification accuracy were assessed. A sampling time of ten days is associated with a high bias of 10% to 47%, which can be reduced to 4.9% to 26.4% if a sampling time of 30 days is used (P < .001). Correct classification of patients archiving glycemic targets can also be improved from 81.5% to 91.9 to 90% to 95.2%. Our results suggest that the proposed 10-14 day CGM sampling time may be associated with a high correlation with three-month CGM. However, these estimates are subject to large intersubject bias, which is clinically relevant. Clinicians and researchers should consider using assessments of longer durations of CGM data if possible, especially when assessing time in hypoglycemia or while testing a new treatment.

Lightweight Multilevel Feature Fusion Network for Hyperspectral Image Classification

Hyperspectral images (HSIs), acquired as a 3D data set, contain spectral and spatial information that is important for ground–object recognition. A 3D convolutional neural network (3DCNN) could therefore be more suitable than a 2D one for extracting multiscale neighborhood information in the spectral and spatial domains simultaneously, if it is not restrained by mass parameters and computation cost. In this paper, we propose a novel lightweight multilevel feature fusion network (LMFN) that can achieve satisfactory HSI classification with fewer parameters and a lower computational burden. The LMFN decouples spectral–spatial feature extraction into two modules: point-wise 3D convolution to learn correlations between adjacent bands with no spatial perception, and depth-wise convolution to obtain local texture features while the spectral receptive field remains unchanged. Then, a target-guided fusion mechanism (TFM) is introduced to achieve multilevel spectral–spatial feature fusion between the two modules. More specifically, multiscale spectral features are endowed with spatial long-range dependency, which is quantified by central target pixel-guided similarity measurement. Subsequently, the results obtained from shallow to deep layers are added, respectively, to the spatial modules, in an orderly manner. The TFM block can enhance adjacent spectral correction and focus on pixels that actively boost the target classification accuracy, while performing multiscale feature fusion. Experimental results across three benchmark HSI data sets indicate that our proposed LMFN has competitive advantages, in terms of both classification accuracy and lightweight deep network architecture engineering. More importantly, compared to state-of-the-art methods, the LMFN presents better robustness and generalization.

Target Classification Method of Tactile Perception Data with Deep Learning.

In order to improve the accuracy of manipulator operation, it is necessary to install a tactile sensor on the manipulator to obtain tactile information and accurately classify a target. However, with the increase in the uncertainty and complexity of tactile sensing data characteristics, and the continuous development of tactile sensors, typical machine-learning algorithms often cannot solve the problem of target classification of pure tactile data. Here, we propose a new model by combining a convolutional neural network and a residual network, named ResNet10-v1. We optimized the convolutional kernel, hyperparameters, and loss function of the model, and further improved the accuracy of target classification through the K-means clustering method. We verified the feasibility and effectiveness of the proposed method through a large number of experiments. We expect to further improve the generalization ability of this method and provide an important reference for the research in the field of tactile perception classification.

Gaussian-guided feature alignment for unsupervised cross-subject adaptation

Human activities recognition (HAR) and human intent recognition (HIR) are important for medical diagnosis and human-robot interaction. HAR and HIR usually rely on the signals of some wearable sensors, such as inertial measurement unit (IMU), but these signals may be user-dependent, which degrades the performance of the recognition algorithm on new subjects. Traditional supervised learning methods require labeling signals and training specific classifiers for each new subject, which is burdensome. To deal with this problem, this paper proposes a novel non-adversarial cross-subject adaptation method called Gaussian-guided feature alignment (GFA). The proposed GFA metric quantifies the discrepancy between the labeled features of source subjects and the unlabeled features of target subjects so that minimizing the GFA metric leads to the alignment of the source and target features. The GFA metric is estimated by calculating the divergence between the feature distribution and Gaussian distribution, as well as the mean squared error of the mean and variance between source and target features. This paper analytically proves the effect of the GFA metric and validates its performance using three public human activity datasets. Experimental results show that the proposed GFA achieves 1% higher target classification accuracy and 0.5% lower variance than state-of-the-art methods in case of cross-subject validation. These results indicate that the proposed GFA is feasible for improving the generalization of the HAR and HIR.

Side-Scan Sonar Image Classification Based on Style Transfer and Pre-Trained Convolutional Neural Networks

Side-scan sonar is widely used in underwater rescue and the detection of undersea targets, such as shipwrecks, aircraft crashes, etc. Automatic object classification plays an important role in the rescue process to reduce the workload of staff and subjective errors caused by visual fatigue. However, the application of automatic object classification in side-scan sonar images is still lacking, which is due to a lack of datasets and the small number of image samples containing specific target objects. Secondly, the real data of side-scan sonar images are unbalanced. Therefore, a side-scan sonar image classification method based on synthetic data and transfer learning is proposed in this paper. In this method, optical images are used as inputs and the style transfer network is employed to simulate the side-scan sonar image to generate “simulated side-scan sonar images”; meanwhile, a convolutional neural network pre-trained on ImageNet is introduced for classification. In this paper, we experimentally demonstrate that the maximum accuracy of target classification is up to 97.32% by fine-tuning the pre-trained convolutional neural network using a training set incorporating “simulated side-scan sonar images”. The results show that the classification accuracy can be effectively improved by combining a pre-trained convolutional neural network and “similar side-scan sonar images”.

Comparison of target classification accuracy according to the aspect angle and the bistatic angle in bistatic sonar

Comparison of target classification accuracy according to the aspect angle and the bistatic angle in bistatic sonar

Target Detection and Classification Performance Enhancement using Super-Resolution Infrared Videos

Long range infrared videos such as the Defense Systems Information Analysis Center (DSIAC) videos usually do not have high resolution. In recent years, there are significant advancement in video super-resolution algorithms. Here, we summarize our study on the use of super-resolution videos for target detection and classification. We observed that super-resolution videos can significantly improve the detection and classification performance. For example, for 3000 m range videos, we were able to improve the average precision of target detection from 11% (without super-resolution) to 44% (with 4x super-resolution) and the overall accuracy of target classification from 10% (without super-resolution) to 44% (with 2x superresolution).

Stress-Induced Effects in Resting EEG Spectra Predict the Performance of SSVEP-Based BCI.

Most research in Brain-Computer-Interfaces (BCI) focuses on technologies to improve accuracy and speed. Little has been done on the effects of subject variability, both across individuals and within the same individual, on BCI performance. For example, stress, arousal, motivation, and fatigue can all affect the electroencephalogram (EEG) signals used by a BCI, which in turn impacts performance. Overcoming the impact of such user variability on BCI performance is an impending and inevitable challenge for routine applications of BCIs in the real world. To systematically explore the factors affecting BCI performance, this study embeds a Steady-State Visually Evoked Potential (SSVEP) based BCI into a "game with a purpose" (GWAP) to obtain data over significant lengths of time, under both high- and low-stress conditions. Ten healthy volunteers played a GWAP that resembles popular match-three games, such as Jewel Quest, Zoo Boom, or Candy Crush. We recorded the target search time, target search accuracy, and EEG signals during gameplay to investigate the impacts of stress on EEG signals and BCI performance. We used Canonical Correlation Analysis (CCA) to determine whether the subject had found and attended to the correct target. The experimental results show that SSVEP target-classification accuracy is reduced by stress. We also found a negative correlation between EEG spectra and the SNR of EEG in the frontal and occipital regions during gameplay, with a larger negative correlation for the high-stress conditions. Furthermore, CCA also showed that when the EEG alpha and theta power increased, the search accuracy decreased, and the spectral amplitude drop was more evident under the high-stress situation. These results provide new, valuable insights into research on how to improve the robustness of BCIs in real-world applications.

Decoding Imagined 3D Arm Movement Trajectories From EEG to Control Two Virtual Arms-A Pilot Study.

Background: Realization of online control of an artificial or virtual arm using information decoded from EEG normally occurs by classifying different activation states or voluntary modulation of the sensorimotor activity linked to different overt actions of the subject. However, using a more natural control scheme, such as decoding the trajectory of imagined 3D arm movements to move a prosthetic, robotic, or virtual arm has been reported in a limited amount of studies, all using offline feed-forward control schemes.Objective: In this study, we report the first attempt to realize online control of two virtual arms generating movements toward three targets/arm in 3D space. The 3D trajectory of imagined arm movements was decoded from power spectral density of mu, low beta, high beta, and low gamma EEG oscillations using multiple linear regression. The analysis was performed on a dataset recorded from three subjects in seven sessions wherein each session comprised three experimental blocks: an offline calibration block and two online feedback blocks. Target classification accuracy using predicted trajectories of the virtual arms was computed and compared with results of a filter-bank common spatial patterns (FBCSP) based multi-class classification method involving mutual information (MI) selection and linear discriminant analysis (LDA) modules.Main Results: Target classification accuracy from predicted trajectory of imagined 3D arm movements in the offline runs for two subjects (mean 45%, std 5%) was significantly higher (p < 0.05) than chance level (33.3%). Nevertheless, the accuracy during real-time control of the virtual arms using the trajectory decoded directly from EEG was in the range of chance level (33.3%). However, the results of two subjects show that false-positive feedback may increase the accuracy in closed-loop. The FBCSP based multi-class classification method distinguished imagined movements of left and right arm with reasonable accuracy for two of the three subjects (mean 70%, std 5% compared to 50% chance level). However, classification of the imagined arm movement toward three targets was not successful with the FBCSP classifier as the achieved accuracy (mean 33%, std 5%) was similar to the chance level (33.3%). Sub-optimal components of the multi-session experimental paradigm were identified, and an improved paradigm proposed.