Brightness Changes Research Articles

Neuromorphic Imaging With Joint Image Deblurring and Event Denoising.

Neuromorphic imaging reacts to per-pixel brightness changes of a dynamic scene with high temporal precision and responds with asynchronous streaming events as a result. It also often supports a simultaneous output of an intensity image. Nevertheless, the raw events typically involve a large amount of noise due to the high sensitivity of the sensor, while capturing fast-moving objects at low frame rates results in blurry images. These deficiencies significantly degrade human observation and machine processing. Fortunately, the two information sources are inherently complementary - events with microsecond-level temporal resolution, which are triggered by the edges of objects recorded in a latent sharp image, can supply rich motion details missing from the blurry one. In this work, we bring the two types of data together and introduce a simple yet effective unifying algorithm to jointly reconstruct blur-free images and noise-robust events in an iterative coarse-to-fine fashion. Specifically, an event-regularized prior offers precise high-frequency structures and dynamic features for blind deblurring, while image gradients serve as a kind of faithful supervision in regulating neuromorphic noise removal. Comprehensively evaluated on real and synthetic samples, such a synergy delivers superior reconstruction quality for both images with severe motion blur and raw event streams with a storm of noise, and also exhibits greater robustness to challenging realistic scenarios such as varying levels of illumination, contrast and motion magnitude. Meanwhile, it can be driven by much fewer events and holds a competitive edge at computational time overhead, rendering itself preferable as available computing resources are limited. Our solution gives impetus to the improvement of both sensing data and paves the way for highly accurate neuromorphic reasoning and analysis.

Intelligent Machinery Fault Diagnosis With Event-Based Camera

Event-based cameras are the emerging bio-inspired technology in vision sensing. Different from the traditional standard cameras, the event-based cameras asynchronously record the brightness change per pixel, and have the great merits of high temporal resolution, low energy consumption, high dynamic range, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">etc</i> . While the event-based cameras have been initially exploited in several common vision-based tasks in the recent years, the investigation on machine condition monitoring problem is quite limited. This paper offers the first attempt in the current literature on exploring the contactless event vision data for machine fault diagnosis. A vibration event representation is proposed to transform the event records into typical data samples, and a deep convolutional neural network model is used for processing the event information. To enhance the model robustness against environmental noisy vision events, an event data augmentation method is proposed to introduce variations of the event patterns. A deep representation clustering method is further proposed to improve the pattern recognition performance with respect to different machine health conditions. Experiments on the event vision-based rotating machine fault diagnosis problem are carried out. It is extensively validated that high fault diagnosis accuracies can be obtained using the vision data from the event-based cameras, which are competitive with the popular accelerometer data. Considering the properties of flexibility, portability and data recognizability, the event-based cameras thus provide a promising new tool for contactless machine health condition monitoring and fault diagnosis.

Association between Gray-Scale Ultrasound Imaging and Serological Creatine Kinase for Quantifying Exercise-Induced Muscle Damage: An Observational Study.

Limited evidence has verified if ultrasound imaging (US) can detect post-exercise muscle damage based on size, shape, and brightness metrics. This study aimed to analyze the correlation between creatine kinase (CK) concentration and (as a biomarker of muscle damage) changes in US gray-scale metrics after an exercise-induced muscle damage protocol. An observational study was conducted at a private university lab located in Madrid. Twenty-five untrained and asymptomatic volunteers were enrolled in this study. Baseline demographic data and body composition metrics were collected. In addition, the rectus femoris US data and CK concentration were assessed at baseline and after inducing muscle damage (24 and 48 h later). After calculating time differences for all the outcomes, the correlation between the changes observed with US and biomarkers was assessed. Significant CK concentration increases were found 24 h (p = 0.003) and 48 h (p < 0.001) after exercise. However, no significant changes in muscle size, shape, or brightness were found in any location (p > 0.05 for all). In addition, no significant associations were found between CK changes and US changes (p > 0.05 for all). Gray-scale US is not a sensitive tool for detecting muscle damage, as a protocol of exercise-induced muscle damage confirmed with CK produced no significant gray-scale US changes after 24 or 48 h. In addition, US and CK changes after 24 and 48 h were not associated with each other.

Grape cluster detection based on spatial-to-depth convolution and attention mechanism

Grape cluster detection is a crucial step in the visual tasks of automated grape harvesting. Background and occlusion lead to difficulty in detecting grape clusters under natural environments. Some improvements have been proposed to solve this issue. Firstly, the public dataset is enriched with the data augmentation methods of random brightness change, image flip left-right, and mosaic to strengthen the model's robustness. Secondly, based on the problem of information loss in grape cluster detection, a plug-and-play module of spatial-to-depth convolution (STD-Conv) is added to enrich grape cluster feature information. The original grape features are further fused by converting the spatial dimension of the input image into a depth dimension. Thirdly, a simple, parameter-free attention mechanism (SimAM) is applied to the backbone to improve the weight of grape targets and suppress background interference weight in feature extraction. Experiments show that combining STD-Conv and SimAM can improve the accuracy of YOLOv4, YOLOv5, and YOLOX. The improved YOLOX model achieves the highest 88.4% mean Average Precision (mAP), 87.8% precision, and 79.5% recall. These findings demonstrate that the enhanced YOLOX model performs well for grape cluster detection. This study's conclusion makes some valuable ideas for automated harvesting into grape or other fruit detection.

In-Situ X-Ray Observation of Bubbles in Porous Transport Layer of PEM Water Electrolysis

As renewable energies become increasingly popular as a countermeasure against global warming, hydrogen is attracting attention as a new energy medium. This is because surplus electricity from renewable energy sources, whose generation is unstable, can be used to electrolyze water (water electrolysis) and convert it into storable hydrogen, which can then be reconverted into electricity by fuel cells and used when needed. By using hydrogen for long-term energy storage and utilizing surplus energy, dependence on fossil fuels can be reduced, resulting in a decrease in greenhouse gas emissions. Water electrolysis technology, particularly Proton Exchange Membrane (PEM) water electrolysis, is currently receiving attention due to its high efficiency, high safety, and a large operational range of current density.In PEM water electrolysis cells, bubbles are produced in the liquid water due to an electrochemical reaction in which water is decomposed into oxygen at the catalyst layer inside the anode. There is a concern that the concentration overvoltage increases as these bubbles remain inside the porous transport layer (PTL), resulting in a decrease in the efficiency of electrolysis. However, the behavior of bubbles during the electrolysis has not been fully clarified, and it is necessary to understand the discharge behavior of bubbles in the PTL. This study aims to clarify the bubble retention behavior in the PTL of PEM water electrolysis cell by in-situ bubble visualization using an X-ray microscopy system. Measurements were taken on 2 types of PTL made from carbon or titanium dioxide with different fiber diameters, pore sizes, and wettability to investigate the effect of PTL structure on bubble accumulation and discharge behaviors. Two-dimensional X-ray images were taken before the start of water electrolysis (when the PTLs were filled with liquid water) and during water electrolysis (when bubbles were generated), and the bubble distribution within the PTL and the channel was quantified from the changes in brightness. By tuning the X-ray transmission thickness of the PTL and X-ray energy, the bubble behavior during the electrolysis were successfully visualized. In the case of the carbon PTL (Fig. 1(a), fiber diameter: 10 μm, pore diameter: 150 μm), bubbles were accumulated under the rib area (Fig. 1(b)), while in the case of the titanium PTL (Fig. 1(c), fiber diameter: 25 μm, pore diameter: 50 μm), bubbles were relatively uniformly distributed (Fig. 1(d)). This result is believed to be due to the titanium's superior wettability, which prevents the accumulation of bubbles by inhibiting their bridging between pores. Furthermore, the titanium PTL is attributed to have a lower gas saturation ratio owing to its smaller pore diameter, which effectively prevents bubble coalescence. These results suggested that the pore size and the wettability of the PTL strongly affect the bubble behavior, and the in-situ observation is quite important to understand it. Figure 1

Brightness and contrast corrections for stereocorrelation: Global and instantaneous formulation with spatial regularization

Stereocorrelation (SC) is a versatile experimental method increasingly used to measure 3D surface displacements in complex mechanical tests. However, convergence issues may arise when gray-level conservation is no longer validated, particularly when local specular reflections are observed. Brightness and contrast corrections (BCCs) are added to the Global Stereocorrelation framework, including different spatial regularization strategies. A challenging experimental test case was chosen for validation purposes where extreme brightness and contrast changes were observed. An impact test on laminated glass was monitored with two high-speed cameras. It is shown that BCCs significantly improve the robustness and convergence rate for both calibration and kinematic measurements. Moreover, the numerical implementation cost into a finite element stereocorrelation code is negligible.

Open-DPSM: An open-source toolkit for modeling pupil size changes to dynamic visual inputs

Pupil size change is a widely adopted, sensitive indicator for sensory and cognitive processes. However, the interpretation of these changes is complicated by the influence of multiple low-level effects, such as brightness or contrast changes, posing challenges to applying pupillometry outside of extremely controlled settings. Building on and extending previous models, we here introduce Open Dynamic Pupil Size Modeling (Open-DPSM), an open-source toolkit to model pupil size changes to dynamically changing visual inputs using a convolution approach. Open-DPSM incorporates three key steps: (1) Modeling pupillary responses to both luminance and contrast changes; (2) Weighing of the distinct contributions of visual events across the visual field on pupil size change; and (3) Incorporating gaze-contingent visual event extraction and modeling. These steps improve the prediction of pupil size changes beyond the here-evaluated benchmarks. Open-DPSM provides Python functions, as well as a graphical user interface (GUI), enabling the extension of its applications to versatile scenarios and adaptations to individualized needs. By obtaining a predicted pupil trace using video and eye-tracking data, users can mitigate the effects of low-level features by subtracting the predicted trace or assess the efficacy of the low-level feature manipulations a priori by comparing estimated traces across conditions.

Plant Disease Detection Using Deep Learning

Agricultural development is a critical strategy for promoting prosperity and addressing the challenge of feeding nearly 10 billion people by 2050. Plant diseases can significantly impact food production, reducing both quantity and diversity. Therefore, early detection of plant diseases through automatic detection methods based on deep learning can improve food production quality and reduce economic losses. While previous models have been implemented for a single type of plant to ensure high accuracy, they require high-quality images for proper classification and are not effective with low-resolution images. To address these limitations, this paper proposes the use of pre-trained model based on convolutional neural networks (CNN) for plant disease detection. The focus is on fine-tuning the hyperparameters of popular pre-trained model such as EfficientNetV2S, to achieve higher accuracy in detecting plant diseases in lower resolution images, crowded and misleading backgrounds, shadows on leaves, different textures, and changes in brightness. The study utilized the Plant Diseases Dataset, which includes infected and uninfected crop leaves comprising 38 classes. In pursuit of improving the adaptability and robustness of our neural networks, we intentionally exposed them to a deliberately noisy training dataset. This strategic move followed the modification of the Plant Diseases Dataset, tailored to better suit the demands of our training process. Our objective was to enhance the network's ability to generalize effectively and perform robustly in real-world scenarios. This approach represents a critical step in our study's overarching goal of advancing plant disease detection, especially in challenging conditions, and underscores the importance of dataset optimization in deep learning applications.

Action Recognition and Benchmark Using Event Cameras.

Recent years have witnessed remarkable achievements in video-based action recognition. Apart from traditional frame-based cameras, event cameras are bio-inspired vision sensors that only record pixel-wise brightness changes rather than the brightness value. However, little effort has been made in event-based action recognition, and large-scale public datasets are also nearly unavailable. In this paper,we present an event-based action recognition framework called EV-ACT. The Learnable Multi-Fused Representation (LMFR) is first proposed to integrate multiple event information in a learnable manner. The LMFR with dual temporal granularity is fed into the event-based slow-fast network for the fusion of appearance and motion features. A spatial-temporal attention mechanism is introduced to further enhance the learning capability of action recognition. To prompt research in this direction, we have collected the largest event-based action recognition benchmark named THUE-ACT-50 and the accompanying THUE-ACT-50-CHL dataset under challenging environments, including a total of over 12,830 recordings from 50 action categories, which is over 4 times the size of the previous largest dataset. Experimental results show that our proposed framework could achieve improvements of over 14.5%, 7.6%, 11.2%, and 7.4% compared to previous works on four benchmarks. We have also deployed our proposed EV-ACT framework on a mobile platform to validate its practicality and efficiency.

Light curves for 10 R Coronae Borealis stars for longer than a century: secular evolution, dip statistics, and a general model for the shape of isolated light-curve dips

ABSTRACT R Coronae Borealis stars (RCBs) are cool supergiants that display non-periodic deep dips in brightness. Recently, a group of ‘Hot RCB stars’ has been discovered to be fast evolving across the HR (Hertzsprung-Russell) diagram, as these stars leave the RCB region, with brightness changes at the rate of ∼1 mag per century. Perhaps cool RCB stars can also be seen evolving, either increasing in temperature as they evolve to become Hot RCB stars, or perhaps increasing in luminosity as the stars arrive at the RCB region. To seek these changes, the only possible method is to extract archival data going back more than a century, looking for the brightness changes associated with the evolution. I have measured and extracted 323 464 mag (mostly from the Harvard plates and from the American Association of Variable Star Observers) for 10 cool RCB stars, all with over a century for the light curves, all consistently calibrated to a modern magnitude system. For times away from any dips, these light curves are flat to within the typical uncertainty of ±0.10 mag per century. That is, I see no significant evolution. I also have collected a large data base of light-curve dips and their properties. From this, the light curves for all the well-observed isolated dips have the same shape, featuring a flat slope for the few days immediately after the minima. Further, I derive a general model for the shape of the light curve for all isolated RCB dips, with a simple equation accurately describing the observed recovery to maximum light.

Rapid identification of boiling crisis with event-based visual streaming analysis

The formation of a stable vapor layer during boiling processes triggers the boiling crisis that can damage the entire system. While there have been efforts to build a predictive model for critical heat flux on live images, unfortunately, images from conventional frame-cameras show limitations for performing real-time digital inference as they contain significant unnecessary information. A new type of imaging technique, neuromorphic event imaging, captures individual pixel-wise brightness changes, thereby recording motion-based phenomena in a sparse and efficient manner. In this study, we develop neural network models capable of predicting the onset of boiling crisis by utilizing neuromorphic-simulated event data. We reveal that event-based models demonstrate their proficiency and time efficiency in learning patterns of the boiling process from relatively small datasets. The event-based models also prove the generalizability to unseen surfaces by showing the potential for transfer learning. The results suggest that an extension of this work may prove effective in developing forecasting models that detect the incipient boiling crisis before it damages boiling surfaces. These findings highlight that event data can be effectively applied to real-time, motion-dependent contexts.

In-situ observation on the δ to γ transformation behavior of carbon steels

In recent in-situ observations on the solidification of thin carbon steel experiments, the primary δ-phase appears to be highly undercooled into the single γ phase region of the phase diagram, where the massive δ to γ transformation occurred without the intervention of the liquid phase following a typical peritectic reaction. Whether such a massive transformation is the main mode, even in bulk steel alloys, is of prime importance because it would drastically change our view on the defect generation mechanism during steel solidification in continuous casting. In this research, the in-situ imaging technique using synchrotron white X-ray beam was conducted to investigate the effects of carbon content and the specimen thickness on the δ to γ transformation behaviors in the carbon steel. The moment of δ to γ transformation is identified through synchronized recording of Laue diffraction patterns and in-situ images. The relative average brightness (RAB) change during solidification was evaluated to analyze the transformation. The results revealed that hypo-peritectic steel grades predominantly underwent a massive transformation, while hyper-peritectic steel grades showed a combination of massive and diffusional transformations. The sudden drop in liquid fraction after massive transformation was detected, while diffusional transformation exhibited a gradual decrease in the present study. The findings demonstrated that specimen thickness influenced the transformation mode in hyper-peritectic steel grades. The study provides insights into the behavior of δ to γ transformation and its implications for carbon steel solidification processes.

The stellar occultation by (319) Leona on 2023 September 13 in preparation for the occultation of Betelgeuse

ABSTRACT On 2023 December 12, the star α Orionis will be occulted by asteroid (319) Leona. This represents an extraordinary and unique opportunity to analyse the brightness distribution of Betelgeuse’s photosphere with extreme angular resolution by studying light curves from different points on Earth and at different wavelengths. Here we present observations of another occultation by asteroid Leona, on 2023 September 13, whose goal was to determine Leona’s projected shape and size in preparation for the December 12th event and its interpretation. The occultation campaign was highly successful with 25 positive detections from 17 different sites and a near miss. The effective diameter in projected area derived from the positive detections was 66 ± 2 km using an elliptical fit to the instantaneous limb. The body is highly elongated, with dimensions of 79.6 ± 2.2 km × 54.8 ± 1.3 km in its long and short axis, respectively, at occultation time. This result, in combination with dense time series photometry of Leona that we recently obtained, together with archival sparse photometry, allowed us to predict the angular size of the asteroid for the Betelgeuse event and to simulate the expected brightness change. Also, an accurate position coming from the occultation is provided, to improve the orbit of Leona.

Neuromorphic imaging and classification with graph learning

Bio-inspired neuromorphic cameras asynchronously record pixel brightness changes and generate sparse event streams. They can capture dynamic scenes with little motion blur and more details in extreme illumination conditions. Due to the multidimensional address-event structure, most existing vision algorithms cannot properly handle asynchronous event streams. While several event representations and processing methods have been developed to address such an issue, they are typically driven by a large number of events, leading to substantial overheads in runtime and memory. In this paper, we propose a new graph representation of the event data and couple it with a Graph Transformer to perform accurate neuromorphic classification. Extensive experiments show that our approach leads to better results and excels at the challenging realistic situations where only a small number of events and limited computational resources are available, paving the way for neuromorphic applications embedded into mobile facilities.

Combustion of polyethylene and polypropylene in the fluidized bed with a variable vertical density profile

A new type of fluidized bed with a variable vertical density profile was developed. A mixture of sand (Ønominal = 100–140 μm, 2.6 g/cm3) and cenospheres (Ønominal = 140–160 μm, 0.85 g/cm3) was used to form a binary fluidized bed. The minimum fluidization velocity of the binary fluidized bed was determined experimentally. Two points of inflection on the fluidization curve were observed. Additionally, a measurement of the pressure as a function of height was used to determine the profile of fluidized bed density. It was proven that the fluidized bed made out of cenospheres and sand can be characterized by a variable vertical density profile. The binary fluidized bed was used for the combustion of polyolefins. The results were compared to combustion in a conventional fluidized bed made out of sand. The surface of the fluidized bed was video recorded and the changes in the image brightness were used to identify diffusion combustion in the freeboard. A Fourier Transform Infrared (FTIR) analyzer with high spectral resolution (1 cm−1) was used to record the IR spectra of flue gases. In-house software was used to deconvolve complex IR spectra. In the process performed in the fluidized bed made out of sand, regardless of temperature, around 30 % of the polymeric carbon left the reactor in the form of soot. The lack of dependence on the temperature results from the fact that in such an organized process combustion did not take place inside the fluidized bed but on its surface. The binary fluidized bed eliminated the problem of the presence of diffusion flames and soot formation. High fuel conversion to CO2 (above 90 %) was observed at temperatures above over 800 °C during polyolefin combustion in the binary fluidized bed.

A Normalized Absolute Values Adaptive Evaluation Function of Image Clarity.

The clarity evaluation function plays a vital role in the autofocus technique. The accuracy and efficiency of the image clarity evaluation function directly affects the accuracy of autofocus and the speed of focusing. However, classical clarity function values are sensitive to changes in background brightness and changes in object contour length. This paper proposes a normalized absolute values adaptive (NAVA) evaluation function of image clarity. It can eliminate the influence of changes in background brightness and the length of the measured object contour on the image clarity function value. To verify the effectiveness of the NAVA function, several experiments were conducted under conditions of virtual master gear images and actual captured images. For actual captured images, the variation of the evaluation results of the NAVA function is far less than the corresponding variation of the classic clarity function. Compared with classical clarity evaluation functions, the NAVA function can provide normalized absolute clarity values. The correlations between the NAVA function results of image clarity and both the contour length and background brightness of the tested object are weak. The use of the NAVA function in automatic and manual focusing systems can further improve focusing efficiency.

A Real-Time Vessel Detection and Tracking System Based on LiDAR.

Vessel detection and tracking is of utmost importance to river traffic. Efficient detection and tracking technology offer an effective solution to address challenges related to river traffic safety and congestion. Traditional image-based object detection and tracking algorithms encounter issues such as target ID switching, difficulties in feature extraction, reduced robustness due to occlusion, target overlap, and changes in brightness and contrast. To detect and track vessels more accurately, a vessel detection and tracking algorithm based on the LiDAR point cloud was proposed. For vessel detection, statistical filtering algorithms were integrated into the Euclidean clustering algorithm to mitigate the effect of ripples on vessel detection. Our detection accuracy of vessels improved by 3.3% to 8.3% compared to three conventional algorithms. For vessel tracking, L-shape fitting of detected vessels can improve the efficiency of tracking, and a simple and efficient tracking algorithm is presented. By comparing three traditional tracking algorithms, an improvement in multiple object tracking accuracy (MOTA) and a reduction in ID switch times and number of missed detections were achieved. The results demonstrate that LiDAR point cloud-based vessel detection can significantly enhance the accuracy of vessel detection and tracking.

Rotation periods and colours of 10-m-scale near-Earth asteroids from CFHT target of opportunity streak photometry

ABSTRACT The rotational properties of ∼10-m-scale asteroids are poorly understood with only a few measurements. Additionally, collisions or thermal recoil can spin their rotations to periods less than a few seconds obfuscating their study due to the observational cadence imposed by the long read-out times of charge-coupled device imagers. We present a method to measure the rotation periods of 10-m-scale asteroids using the target of opportunity capability of the Canada–France–Hawaii Telescope (CFHT) and its MegaCam imager by intentionally streaking their detections in single exposures when they are at their brightest. Periodic changes in brightness as small as ∼0.05 mag along the streak can be measured as short as a few seconds. Additionally, the streak photometry is taken in multiple g, r, and i filter exposures enabling the measurement of asteroid colours. The streak photometry method was tested on CFHT observations of three 10-m-scale asteroids, 2016 GE1, 2016 CG18, and 2016 EV84. Our three targets are among the smallest known asteroids with measured rotation periods/colours having some of the shortest known rotation periods. We compare our rotation period and taxonomic results with independent data from the literature and discuss applications of the method to future small asteroid observations.

Study on the Rapid Degradation Performance of Salix/Wheat Straw Fiber Degradable Film

The preparation of biodegradable mulch film to replace non-degradable mulch film is of great significance for reducing the harm of non-degradable agricultural mulch film to the environment. However, there are few studies on the degradation performance and degradation mechanisms of degradable cellulose mulch. Therefore, the wet papermaking process was adopted in this work. Salix fiber and wheat straw fiber were used as raw materials. A Salix/wheat straw fiber degradable film was prepared by adding cationic polyacrylamide, alkyl ketene dimer, and paraffin emulsion. The degradation process of cellulose film was studied using a UV degradation test and an acid-base degradation test system. The results showed that after 40 days of UV degradation, the degradation rate of Salix/wheat straw fiber degradable film could reach 6.66%. The tensile strength could still maintain 2.878 KN/m. The results of the brightness change index (ΔL) and color overall change index (ΔE) showed that the surface of the Salix/wheat straw fiber degradable film had been successfully partially degraded. After 4 days of alkaline degradation, the degradation rate could reach 11.89%. After 4 days of acid degradation, the degradation rate could reach 14.64%. At the same time, the specific degradation process of Salix/wheat straw fiber degradable film was further studied by infrared spectroscopy and scanning electron microscopy. This work provides a new method for the study of agricultural degradable cellulose mulch, which is of great significance for the future development of agricultural mulch.

A temporal densely connected recurrent network for event-based human pose estimation

Event camera is an emerging bio-inspired vision sensors that report per-pixel brightness changes asynchronously. It holds noticeable advantage of high dynamic range, high speed response, and low power budget that enable it to best capture local motions in uncontrolled environments. This motivates us to unlock the potential of event cameras for human pose estimation, as the human pose estimation with event cameras is rarely explored. Due to the novel paradigm shift from conventional frame-based cameras, however, event signals in a time interval contain very limited information, as event cameras can only capture the moving body parts and ignores those static body parts, resulting in some parts to be incomplete or even disappeared in the time interval. This paper proposes a novel densely connected recurrent architecture to address the problem of incomplete information. By this recurrent architecture, we can explicitly model not only the sequential but also non-sequential geometric consistency across time steps to accumulate information from previous frames to recover the entire human bodies, achieving a stable and accurate human pose estimation from event data. Moreover, to better evaluate our model, we collect a large-scale multimodal event-based dataset that comes with human pose annotations, which is by far the most challenging one to the best of our knowledge. The experimental results on two public datasets and our own dataset demonstrate the effectiveness and strength of our approach. Code is available online for facilitating the future research.