Achievable Detection Performance Research Articles

Transfer learning with generative models for object detection on limited datasets

Abstract The availability of data is limited in some fields, especially for object detection tasks, where it is necessary to have correctly labeled bounding boxes around each object. A notable example of such data scarcity is found in the domain of marine biology, where it is useful to develop methods to automatically detect submarine species for environmental monitoring. To address this data limitation, the state-of-the-art machine learning strategies employ two main approaches. The first involves pretraining models on existing datasets before generalizing to the specific domain of interest. The second strategy is to create synthetic datasets specifically tailored to the target domain using methods like copy-paste techniques or ad-hoc simulators. The first strategy often faces a significant domain shift, while the second demands custom solutions crafted for the specific task. In response to these challenges, here we propose a transfer learning framework that is valid for a generic scenario. In this framework, generated images help to improve the performances of an object detector in a few-real data regime. This is achieved through a diffusion-based generative model that was pretrained on large generic datasets. With respect to the state-of-the-art, we find that it is not necessary to fine tune the generative model on the specific domain of interest. We believe that this is an important advance because it mitigates the labor-intensive task of manual labeling the images in object detection tasks. We validate our approach focusing on fishes in an underwater environment, and on the more common domain of cars in an urban setting. Our method achieves detection performance comparable to models trained on thousands of images, using only a few hundreds of input data. Our results pave the way for new generative AI-based protocols for machine learning applications in various domains, for instance ranging from geophysics to biology and medicine.

YOLIC: An efficient method for object localization and classification on edge devices

In the realm of Tiny AI, we introduce “You Only Look at Interested Cells” (YOLIC), an efficient method for object localization and classification on edge devices. Through seamlessly blending the strengths of semantic segmentation and object detection, YOLIC provides improved computational efficiency and precision compared to traditional methods. By adopting Cells of Interest for classification instead of individual pixels, YOLIC encapsulates relevant information, reduces computational load, and enables rough object shape inference. Importantly, the need for bounding box regression is obviated, as YOLIC capitalizes on the predetermined cell configuration that provides information about potential object location, size, and shape. To tackle the issue of single-label classification limitations, a multi-label classification approach is applied to each cell for effectively recognizing overlapping or closely situated objects. This paper presents extensive experiments on multiple datasets to demonstrate that YOLIC achieves detection performance comparable to the state-of-the-art YOLO algorithms while surpassing in speed, exceeding 30fps on a Raspberry Pi 4B CPU.

Semi-supervised Open-World Object Detection

Conventional open-world object detection (OWOD) problem setting first distinguishes known and unknown classes and then later incrementally learns the unknown objects when introduced with labels in the subsequent tasks. However, the current OWOD formulation heavily relies on the external human oracle for knowledge input during the incremental learning stages. Such reliance on run-time makes this formulation less realistic in a real-world deployment. To address this, we introduce a more realistic formulation, named semi-supervised open-world detection (SS-OWOD), that reduces the annotation cost by casting the incremental learning stages of OWOD in a semi-supervised manner. We demonstrate that the performance of the state-of-the-art OWOD detector dramatically deteriorates in the proposed SS-OWOD setting. Therefore, we introduce a novel SS-OWOD detector, named SS-OWFormer, that utilizes a feature-alignment scheme to better align the object query representations between the original and augmented images to leverage the large unlabeled and few labeled data. We further introduce a pseudo-labeling scheme for unknown detection that exploits the inherent capability of decoder object queries to capture object-specific information. On the COCO dataset, our SS-OWFormer using only 50% of the labeled data achieves detection performance that is on par with the state-of-the-art (SOTA) OWOD detector using all the 100% of labeled data. Further, our SS-OWFormer achieves an absolute gain of 4.8% in unknown recall over the SOTA OWOD detector. Lastly, we demonstrate the effectiveness of our SS-OWOD problem setting and approach for remote sensing object detection, proposing carefully curated splits and baseline performance evaluations. Our experiments on 4 datasets including MS COCO, PASCAL, Objects365 and DOTA demonstrate the effectiveness of our approach. Our source code, models and splits are available here https://github.com/sahalshajim/SS-OWFormer

SSPT-bpMRI: A Self-supervised Pre-training Scheme for Improving Prostate Cancer Detection and Diagnosis in Bi-parametric MRI.

Prostate cancer (PCa) is one of the most prevalent cancers in men. Early diagnosis plays a pivotal role in reducing the mortality rate from clinically significant PCa (csPCa). In recent years, bi-parametric magnetic resonance imaging (bpMRI) has attracted great attention for the detection and diagnosis of csPCa. bpMRI is able to overcome some limitations of multi-parametric MRI (mpMRI) such as the use of contrast agents, the time-consuming for imaging and the costs, and achieve detection performance comparable to mpMRI. However, inter-reader agreements are currently low for prostate MRI. Advancements in artificial intelligence (AI) have propelled the development of deep learning (DL)-based computer-aided detection and diagnosis system (CAD). However, most of the existing DL models developed for csPCa identification are restricted by the scale of data and the scarcity in labels. In this paper, we propose a self-supervised pre-training scheme named SSPT-bpMRI with an image restoration pretext task integrating four different image transformations to improve the performance of DL algorithms. Specially, we explored the potential value of the self-supervised pre-training in fully supervised and weakly supervised situations. Experiments on the publicly available PI-CAI dataset demonstrate that our model outperforms the fully supervised or weakly supervised model alone.

High mobility enabled spatial and media‐based modulated orthogonal frequency division multiplexing systems for beyond 5G wireless communications

SummaryThis paper introduces the concept of spatial and media‐based modulated (SMBM) orthogonal frequency division multiplexing (OFDM) as a potential candidate for highly mobile next generation beyond 5G (B5G) wireless communications. The proposed SMBM‐OFDM technique utilizes not only the transmit antenna and channel state indices but also OFDM subcarriers to improve the system performance under high mobility. In addition, this study sheds light on challenging fast time‐varying channel estimation problem of MBM‐based systems by using the linear minimum mean square error (LMMSE) approach due to its optimality to investigate the achievable system performance with the aid of basis expansion modeling. The minimum lower bound on the channel estimation error (Bayesian Cramer–Rao bound) is derived theoretically and shown to be attainable by the considered LMMSE estimator. Moreover, symbol detection performance is provided for different modulation types and higher mobile velocities. Simulation results demonstrate that SMBM‐OFDM system under high mobility is able to provide around 12‐dB performance gains in terms of both channel estimation and symbol detection error compared to conventional spatial modulation (SM)‐OFDM systems without MBM. The presented framework is important due to addressing the high mobility support of SMBM‐OFDM systems for B5G wireless communications in terms of achievable channel estimation and data detection performance.

Scale-adaptive local intentional surface feature detection

In this paper, we propose a mesh-based feature detection scheme that focuses on surface features. A class of features of key interest is intentional structures that act as fiducials and that, for instance, can assist in shape retrieval and distortion measurement. We introduce a tunable two-scale depth measurement scheme to quantify the displacement of a vertex from the local surface, which can be a strong indicator of features. We print and scan 3D models with fiducial features appearing across the surface to demonstrate the high fidelity and accuracy of the proposed feature detection scheme. The method outperforms existing 3D feature detection schemes on CAD models and 3D scans alike. We also discuss applications of data embedding enabled by the achievable detection performance.

Spatial Diversity in Radar Detection via Active Reconfigurable Intelligent Surfaces

Active reconfigurable intelligent surfaces (RISs) are a novel and promising technology that allows controlling the radio propagation environment while compensating for the product path loss along the RIS-assisted path. In this letter, we consider the classical radar detection problem and propose to use an active RIS to get a second independent look at a prospective target illuminated by the radar transmitter. At the design stage, we select the power emitted by the radar, the number of RIS elements, and their amplification factor in order to maximize the detection probability for a fixed probability of false alarm and a common (among radar and RIS) power budget. An illustrative example is provided to assess the achievable detection performance, also in comparison with that of a radar operating alone or with the help of a passive RIS.

Remote interference management in 5G new radio: methods and performance

In time division duplexing based mobile networks, under certain atmospheric ducting conditions, the uplink reception may be interfered by the downlink transmissions of remote base-stations (gNBs) located hundreds of kilometers away. This paper addresses such remote interference problem in a 5G new radio (NR) macro deployment context. Specifically, two potential reference signal (RS) designs for remote interference management (RIM) are described. The first signal structure, denoted as the one OFDM symbol (1OS) based RIM-RS, is building on the channel state information reference signals of 5G NR. The second candidate is referred to as the two OFDM symbol based RIM-RS design, which builds on the design principles of LTE RIM-RS. The achievable detection performance is evaluated by introducing enhanced receiver algorithms together with three feasible propagation delay based gNB grouping and corresponding RIM-RS transmissions schemes. The performance results in terms of the receiver processing gain highlight that the improved detection algorithm assures sufficient performance to detect the remote interference for both RIM-RSs with all evaluated frequency domain comb-like patterns. The benefit of grouping corresponding RIM-RS transmissions from gNBs located on the same area is greater when using same frequency domain resources per transmitted sequence in practical interference scenarios. Furthermore, applying a common base sequence for all gNBs within a group allows to identify the group based on detected sequence and enables adaptive RIM mitigation schemes. On the other hand, it is shown that the 1OS RIM-RS provides smaller overhead and can be frequency multiplexed with the physical downlink shared channel, which opens up the possibility of using gNB group wise 1OS RIM-RS also for UE interference measurements.

Soft Output Signal Detection for Massive MIMO Systems Based on Chebyshev Trace Iteration

Minimum Mean Square Error (MMSE) algorithm can achieve near-optimal detection performance for multi-user massive Multiple Input Multiple Output (MIMO) systems. However, the calculation of the high-dimensional matrix inversion required in MMSE algorithm causes excessively high computational complexity, which makes it difficult to implement quickly and effectively in practical applications. At the same time, for the Higher Quadrature Amplitude Modulation (HQAM), the direct use of hard decision to realize the symbol-to-bit demapper will result in an obvious performance loss. Therefore, a low-complexity soft output signal detection algorithm based on Chebyshev-Trace Iteration (CTI) is proposed for Gray-coded HQAM in this paper. The algorithm not only effectively avoids the calculation of the high dimensional matrix inversion, but also gives a simplified calculation method with the trident list searching to compute the bit Log-Likelihood Ratio (LLR) by using the bit flip property of Gray-coded modulation and binary tree architecture. The simulation results show that the proposed soft output detection algorithm needs at most 3 iterations to converge and achieve detection performance close to MMSE, which achieves a good tradeoff between the complexity and the detection performance.

Exploring a Multimodal Mixture-Of-YOLOs Framework for Advanced Real-Time Object Detection

To construct a safe and sound autonomous driving system, object detection is essential, and research on fusion of sensors is being actively conducted to increase the detection rate of objects in a dynamic environment in which safety must be secured. Recently, considerable performance improvements in object detection have been achieved with the advent of the convolutional neural network (CNN) structure. In particular, the YOLO (You Only Look Once) architecture, which is suitable for real-time object detection by simultaneously predicting and classifying bounding boxes of objects, is receiving great attention. However, securing the robustness of object detection systems in various environments still remains a challenge. In this paper, we propose a weighted mean-based adaptive object detection strategy that enhances detection performance through convergence of individual object detection results based on an RGB camera and a LiDAR (Light Detection and Ranging) for autonomous driving. The proposed system utilizes the YOLO framework to perform object detection independently based on image data and point cloud data (PCD). Each detection result is united to reduce the number of objects not detected at the decision level by the weighted mean scheme. To evaluate the performance of the proposed object detection system, tests on vehicles and pedestrians were carried out using the KITTI Benchmark Suite. Test results demonstrated that the proposed strategy can achieve detection performance with a higher mean average precision (mAP) for targeted objects than an RGB camera and is also robust against external environmental changes.

Adaptive Bayesian Detection for MIMO Radar in Gaussian Clutter

For collocated Multiple-Input Multiple-Output (MIMO) radar, we investigate the target detection problem in Gaussian clutter with an unknown but random covariance matrix. An inverse complex Wishart distribution is chosen as prior knowledge for the random covariance matrix. We propose two detectors in the Bayesian framework based on the criteria of the Generalized Likelihood Ratio Test. The two main advantages of the proposed Bayesian detectors are as follows: (1) no training data are required; and (2) a prior knowledge about the clutter is incorporated in the decision rules to achieve detection performance gains. Numerical simulations show that the proposed Bayesian detectors outperform the current commonly used non-Bayesian counterparts, particularly when the sample number of the transmitted waveform is small. In addition, the performance of the proposed detector will decline in parameter mismatched situation.

A low-complexity photoplethysmographic systolic peak detector for compressed sensed data

Objective: Recent advances in wearable technologies and signal processing have made it possible to perform health monitoring during everyday life activities. Despite the fact that new technologies allow the storage of large volumes of data on small devices, limitations remain when data have to be transmitted or processed with devices with both energy and computational constraints. Approach: This work focuses on the implementation and validation of a photoplethysmogram (PPG) low-complexity analysis method for sensors that acquire a compressed PPG signal through compressive sensing (CS) and allows for the accurate detection of the PPG systolic peak in the compressed domain. Three public datasets were used consisting of a total of about 52 h of PPG signals from 600 patients with normal and abnormal rhythms. Peaks were manually annotated by experts or derived from the annotated synchronized ECG. Main results: The proposed method achieved a pooled average F1 measure on the three datasets of 91% 8% for a 5% compression ratio (CR), 89% 10% for CR = 70% and 82% 12% for CR of 90%. The pooled average F1 measure on the original uncompressed data using an offline open source peak detector is F1 = 91% 11%. The proposed method is up to ∼100 times faster with respect to methods using decompression followed by peak detection. Significance: Results demonstrate that it is possible to achieve detection performance, in terms of the F1 measure, comparable with those obtained on the original uncompressed and filtered signal, making the proposed approach appropriate for real-time wearable systems with energy and computation constraints.

Detection of Sparse Stochastic Signals With Quantized Measurements in Sensor Networks

In this paper, we consider the problem of detection of sparse stochastic signals with quantized measurements in sensor networks. The observed sparse signals are assumed to follow the Bernoulli–Gaussian distribution. Due to the limited bandwidth in sensor networks, the local sensors are required to send quantized measurements to the fusion center. First, we propose a detector using the locally most powerful test (LMPT) strategy, called the quantized LMPT detector, for the problem of distributed detection of sparse signals with quantized measurements. Then, the local quantizers are designed to guarantee the near optimal detection performance of the proposed quantized LMPT detector. When the designed quantization thresholds are applied at the local sensors, we show that 1) the proposed 1-bit LMPT detector with 3.3 L sensors achieves approximately the same detection performance as the clairvoyant LMPT detector with L sensors; 2) the proposed LMPT detector with 3-bit measurements can achieve detection performance comparable to the clairvoyant LMPT detector. Simulation results demonstrate the performance of the proposed quantized LMPT detector and corroborate our theoretical analysis.

Bayesian Detection for MIMO Radar in Gaussian Clutter

For colocated multiple-input multiple-output (MIMO) radar, we investigate the target detection problem in Gaussian clutter with unknown covariance matrix with known inverse complex Wishart distribution as its prior probability density function. We propose three detectors in the Bayesian framework according to the criteria of the generalized likelihood ratio test, Rao test, and Wald test. The two main advantages of the proposed Bayesian detectors are as follows: first, no training data are required; and second, a priori knowledge about the clutter is incorporated in the decision rules to achieve detection performance gains. Numerical simulations show that the proposed Bayesian detectors outperform their non-Bayesian counterparts, especially when the sample number of the transmitted waveform is small. In addition, the proposed Bayesian Wald test is the most robust against the mismatch in the receive steering vector, and the proposed Bayesian Rao test exhibits the strongest rejection capability of mismatched signals.

Temporal Action Detection in Untrimmed Videos from Fine to Coarse Granularity

Temporal action detection in long, untrimmed videos is an important yet challenging task that requires not only recognizing the categories of actions in videos, but also localizing the start and end times of each action. Recent years, artificial neural networks, such as Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) improve the performance significantly in various computer vision tasks, including action detection. In this paper, we make the most of different granular classifiers and propose to detect action from fine to coarse granularity, which is also in line with the people’s detection habits. Our action detection method is built in the ‘proposal then classification’ framework. We employ several neural network architectures as deep information extractor and segment-level (fine granular) and window-level (coarse granular) classifiers. Each of the proposal and classification steps is executed from the segment to window level. The experimental results show that our method not only achieves detection performance that is comparable to that of state-of-the-art methods, but also has a relatively balanced performance for different action categories.

Secrecy Constrained Distributed Detection in Sensor Networks

The data collected by sensor networks often contain sensitive information and care must be taken to prevent that information from being leaked to malicious third parties, e.g., eavesdroppers. Under both the Neyman–Pearson and Bayesian frameworks, we investigate the strategy of defending against an informed and greedy eavesdropper who has access to all the sensors’ outputs via imperfect communication channels. Meanwhile, the legitimate user, e.g., fusion center, is guaranteed to achieve its desired detection performance. Under the Neyman–Pearson framework, we propose a novel approach for analyzing the performance tradeoff, using the maximum achievable detection performance ratio between the fusion center and eavesdropper. Under the Bayesian framework, we derive the asymptotic error exponent, and show that the detectability of a given eavesdropper (Eve) can be limited to her prior information, in other words, Eve's observations do not improve her decision-making ability. Furthermore, we show that as the number of sensors goes to infinity, both asymptotic perfect secrecy and asymptotic perfect detection can be achieved under both frameworks, given noiseless communication channels to the fusion center.

Detection Performance of a Forward Scatter Radar Using a Crystal Video Detector

The forward scatter radar (FSR) configuration is especially appealing for the detection of low-observable targets since it provides well-known properties for the radar cross section, which includes enhancements with respect to the monostatic and/or moderate bistatic configurations, as well as robustness to the target material and detailed geometrical characteristics. Due to the separation of a transmitter and a receiver, it is customary to use a detection scheme based on a square-law envelope detector followed by an appropriate matched filter, which we address as a crystal video detector (CVD) based on traditional terminology. This paper provides an accurate analytical expression for the detection performance of the FSR target detection using the CVD in terms of both the probability of false alarm $(P_{{\rm{fa}}})$ and the probability of detection $(P_{{\rm{d}}})$ in order to support performance prediction and FSR system design. The derived expressions are validated by comparison to Monte Carlo simulations under two different geometrical scenarios. Moreover, the comparison of the CVD performance to an ideal optimum detector (upper bound for the FSR achievable detection performance) shows the geometrical conditions, where the CVD suffers significant detection losses so that alternative detectors should be investigated. Finally, to remove the need to operate the CVD with a fixed detection threshold, two fully adaptive detectors are derived, based on the structure of the CVD scheme, which are shown to provide a constant false alarm rate (CFAR). The performance of these CFAR detectors in terms of $P_{{\rm{fa}}}$ and $P_{{\rm{d}}}$ are provided in closed form and validated through Monte Carlo simulations, showing quite small losses with respect to the fixed-threshold CVD. The application to real data acquired by a passive FSR based on frequency-modulated signals is used to show the practical effectiveness and consistency.

Hidden Markov Model Based Signal Characterization for Weak Light Communication

Weak illumination by light emitting diode (LED) can be realized by either peak-power transmission at a low duty cycle or low-power transmission at a regular duty cycle. In this paper, we address the signal characterization and transmission rate of visible light communication for the latter case. We first characterize the LED nonlinear effect via a state transition model based on the experimental measurements. Then, we adopt hidden Markov model (HMM) to characterize the LED transmission under weak illumination. Based on the HMM, we propose a Monte Carlo method to compute the achievable transmission rate and apply Viterbi algorithm to detect the signals. The achievable transmission rates and detection performance of the HMM-based model are evaluated by simulation results. Moreover, both numerical and experimental results show lower bit error rate of the Viterbi algorithm compared with the per symbol detection.

Massive MIMO for Distributed Detection With Transceiver Impairments

This paper investigates an issue of massive MIMO-based distributed detection that considers transceiver hardware impairments at both a massive-antenna fusion center (FC) and multiple single-antenna sensors. First, we derive closed-form expressions of the probability of detection and the probability of false alarm, and show that hardware impairments create finite ceilings on the achievable detection performance. Then, we formulate a constrained optimization problem as sum of linear ratios programming to maximize the detection probability. By exploiting the inherent problem structures, we further develop an algorithm based on the alternating direction method of multipliers. Extensive simulations demonstrate that the nonideal hardware has a fundamental impact on the distributed detection performance. More specifically, in the limit of an infinite number of antennas and infinite sensor reporting power budget, the effects of FC impairment and FC receiver noise vanish, while the sensor impairment dominates the achievable distributed detection performance.

Random-vibration-based damage detection and precise localization on a lab–scale aircraft stabilizer structure via the Generalized Functional Model Based Method

The problem of random-vibration-based damage diagnosis, including detection and localization, for a lab–scale aircraft stabilizer structure is addressed. Diagnosis is based on the recently introduced Generalized Functional Model Based Method, which utilizes a simple data-based model – instead of large-scale finite element type models – in the inspection phase and is capable of estimating the damage coordinates and their uncertainty region. The focus of the work is on assessing the achievable detection performance and localization accuracy under constraints on the number of deployed sensors and the excitation bandwidth, both reflecting practical limitations relating to potential in-flight implementation. The damage scenarios simulate local stiffness reduction via the addition of a small mass at any point on the structure. While the method has been recently employed with structures consisting of one-dimensional elements, it is presently and for the first time challenged with the considerably more complex problem of damage localization on a two-dimensional structural element. The method’s details, including the transformation of functional boundary constraints into simple ones within the context of the associated optimization problem, are illustrated. The results obtained with a large number of experiments, using only two vibration sensors and limited, low-frequency signal bandwidth, demonstrate excellent damage detection and remarkably accurate localization, thus indicating the method’s high potential for effective damage diagnosis.