Deep Learning

In Taiwan, there are almost 96 million people look for medical treatment due to diabetes. The project successfully used the infrared light measurement system to observe the differences between the bilateral feet,collocating with the Chaos Synchronization to measure the differences between different signals, to decrease the probability of the attendance as the ultimate objective .

"1. Automatically generate Verilog code tools based on the hardware architecture of convolutional neural networks: 4 different hardware architectures (output stationary, weight stationary, Tree architecture, NVDLA) can be generated for the currently more commonly used DNN networks. 2. Visual performance index analysis tool: According to the selected DNN model,the choice of hardware architecture specifications, analyze the performance index."

"For video anomaly detection, we apply pretrained models to obtain the foregroundthe optical flow as ground truth. Then our model estimates the information by taking only a single frame as input. For human behaviors, we take the human poses as inputuse a GCN-based model to predict the future poses. Both the anomaly scores of these two works are given by the error of the estimation. For defect detection, our model takes patches of the image as inputlearns to extract features. The anomaly score of each patch is given by the distance between the patchall the training patches."

We provides a wearable device for the visually impaired to walk outdoors. By the deep learning network, the system can recognizeguide the visually impaired to walk on safe areas such as sidewalkscrosswalk. In addition, it can recognize the types of common obstaclesguide the visually impaired to avoid it in advance. Finally, we can convert the Google Maps route into easy-to-understand voice prompts instruction to guide the visually impaired to move in the right direction.

Super fast terapixel server allows multiple users to viewanalyze gigapixelterapixel images with mobile devices, such as iPhones, iPads, Android mobiles,laptops at the same time. Users could annotateanalyze the data using built-in tools such as 3d effect, color deconvolution, color quantification, nuclei/object detection, lengtharea measurement,quantitative reporting system.

Our system is achieved by integrating diffractiondeep learning methods, which allows high-throughput lensless imaging system to display wide rangehigh-resolution images. In addition, the developed extraction chipsprototype were verified to accurately detect the numberproportion of blood cells under low blood demand. Our system shows great potential in point-of-care blood cells monitoring for cancer patients that could reduce infection riskmortality rate, increase efficacy of chemotherapysupport precision medicine.

"Four automatic rumor detection models based on artificial intelligencenatural language processing. Rumor Detection on Twitter Using Multiloss Hierarchical BiLSTM with an Attenuation Factor Exploiting Microblog Conversation Structures to Detect Rumors Birds of a Feather Rumor Together? Exploring HomogeneityConversation Structure in Social Media for Rumor Detection Meet The Truth: Leverage Objective FactsSubjective Views for Interpretable Rumor Detection Experimental results show that these four models have greatly improved the accuracy of rumor detection, surpassing other SOTA models."

We invented a non-destructive terahertz (THz) deep-learning computed tomography system based on time-domain spectroscopy. In the method, THz time-domain signals are profiled. Multiple features are retrieved from those profiles by a trained modeltransformed to the spatial domain to reconstruct a cross-sectional tomographic image. We have also invented a 3D THz tomographic system based on multi-scale spatio-spectral feature fusion in a multi-scale manner. We believe our work will stimulate further applicable research of THz tomographic imaging with advanced computer vision techniques.

This theme is designed to implement the robot＇s appearancepractical functions. Apply PSPNet to detect the walkable planeYolo to detect obstacles, so that the robot has the autonomous obstacle avoidance function, informing more information about the environmental obstacles around the visually impaired,apply CNN to locate indoor position with self-built indoor database.

With the astronomically growing trade flows, customs administrations need effective and explainable methods to detect suspicious transactions. This project presents a novel artificial intelligence-based model named DATE that ranks trade flows in the order of fraud risk and to maximize customs revenue. We confirm the superiority of DATE over state-of-the-art AI models, with a remarkable precision of 92.7% on illegal cases and a recall of 49.3% on revenue after inspecting only 1% of all trade flows. Predictions of DATE are also interpretable from the attention mechanism. We are deploying DATE in Nigeria and Malawi Customs Services, in collaboration with the World Customs Organization (WCO). DATE has been published in ACM KDD 2020, which is an AI top conference.

This system provides multiple functions for digital slides, allowing users to complete all tasks without switching between different software systems. The functions include, Project management, Image management, Annotation, Zoon In / Zoom Out view, Navigation view, Category shortcuts, Authority management,Annotation overview.

The sweet pepper cultivationplant protection robot developed includes three parts: the intelligent image identificationpositioning system of sweet pepper diseasesinsect pests is composed of YOLO4 deep learning modelimage 3D vision integration. It can perform disease detectionspatial positioning,guide the robotic arm spraying device to accurately spray the diseased. The automatic walking system can use GPS to plan the movement path of the plant protection robot in the field, search for plants infected with pestsdiseases,perform the tasks of pest control.

"We propose a system that use the most popular deep learning technology to analyze panorama videoimage to extract the 3D layout of the panorama image to create a virtual scene. This 3D layout structure allows the system to add 3D model to the original panorama video, making it more editablecapable of synthesize virtual objects to the real world of panorama videos. Our system uses the 3D layout to create low-poly mesh with almost the same quality of the real scenery. This allow users to browse the 3D result on website with low latency."

The system, “Occluded Resistant Face DetectionRecognition System”, contains different scale detectors for calculating face locations. Additive angular margin loss is added into the training phase for achieving high efficiencyaccuracy. The system can achieves 80 accuracy under the 50 face occluded.

This study built an automatic insect pest image identification system based on tiny Yolov3 deep learning model. By optimizing the tiny Yolov3 detection model, images of insect pests on scanned sticky paper can be automatically identified. The system achieves a testing accuracy of 0.93, 0.90 for whitefliesthrips respectively.

We propose a spatial-temporal inference model, which uses a large amount of spatialtemporal data in the city to help governmentsenterprises predict future long-short-term important urban indicator values, such as traffic flow, human mobility, pollution level, number of criminal caseseven commercial profitability. The SIM model exploits IoT to integrate multiple real-time geospatial big data, including population, the flow of people, geographical data, Traffic,real-time sensor values. SIM can make effective predictionsprovide explainability for making decisions.

This study explores the energy-saving efficiency of a complex water-cooled chiller system composed of multiple chillers, cooling towerswater pumps connected by circulation pipelines. The deep learning modeling method driven by big data provides optimized parameters of the water-cooled chiller system to reduce the power consumption of the chiller system. Experiments show that this technology can increase the power saving efficiency of AU Optronics Longtan L5B factory by 2.42. This method has been implemented in many factories of AUO.

A flexible PPG sensor patch for measuring pulsations of blood vessels estimating ambulatory blood pressure monitoring (ABPM). The measured quality PPG signals are analyzed to estimate blood pressure (BP), heart rate (HR), blood oxygen (SpO2)atrial fibrillation (A-Fib)avoid motion artifacts. In this way, with subject’s physiological condition apprehended accurately.

In order to develop an AI model that can accuratelycompletely segment coronary arteries, our team, the TW-CVAI, has established a training dataset composed of strictly verified annotations of coronary lumen boundaries in coronary CT angiography (CCTA). We designed a deep learning model, two-channel 3D-UNet, with a priori prerequisite (vesselness prior) to facilitate identification of vascular structures. The final model, the TaiCAD-Net, greatly shortens the CCTA interpretation time from 6 hours to 10 minutes, with the overall segmentation accuracy of 86 by Dice similarity coefficient.

"1. Use PE Cluster to make the calculation reconfigurable,can be modularized into 36X PEs, supporting up to 2160PEs 2. It can be configured as an independent IP with high computing throughputlow power consumption. The computing performance can reach 864 GOPS@200MHz when 2160PEs are employed."

We develop an asynchronous framework across userkernel spaces for deep learning applications on the improvement of Wi-Fi performanceimplement it in the driver of commodity Intel Wi-Fi cards. Under this framework, we apply deep reinforcement learning to developing an intelligent rate adaptation (RA) algorithm (DRL-RA), which can achieve the highest throughput in varying channel conditions given many rate options of current Wi-Fi technologies. Its on-learning capability can learn how to efficiently approach the best rate from the experiences of its common usage patternenvironment.

We develop an image-based sim-to-real transfer technique for deep reinforcement learning. First, we train a teacher model to move along a near optimal path. We then use this model to teach a student model the correct actions along with randomization. The technique bridges the sim-to-real gap, improving the driving speedrobustness of the simulator-trained student model in the real world.

Our technique is a liver tumor detection method using machine learning which is the most popular AI technique in these years. the high-magnification image provides more details, but the view is smaller in contrast, a low-magnification image has the larger view but less details. Thus, our method combines the two magnificationsyields more accurate tumor detection results.

Due to the fast development of 5G networks, it is critical to identify users service types to allocate resources intelligently. Our technology focuses on users mobility type identification by extracting practical features from users cellular information. We proposed a system architecturehave collected 700-hour data with 150 GB. By using our dataother datasets in the world, we show that our technology can achieve 95 accuracy,reduce 16 energy consumption compared to traditional methods.

Multi-axis force/torque sensors are key components of precision machinesrobot arms. Development of the multi-axis force/torque sensor is a popular research topic in Taiwan. This research developed design methods, calibration machinescalibration algorithms for multi-axis force/torque sensors. These efforts contribute the enhancement of precision of 0.20 (1.25 times than commercialize pro

Deep-Learning-Based Pupil Center DetectionTracking Technology for Visible-Light Wearable Gaze Tracking Devices

The DNN models of this technology include a LithoNet, an OPCNet,a layout novelty detection network. LithoNet is a learning-based pre-simulation model for layout-to-SEM contour prediction,OPCnet is a dual network of LithoNet for photomask optimization. Integrated with a well-trained LithoNet, our layout novelty detection network, consisting of a self-attention guided LithoNetan autoencoder, can check if there are layout patterns easily resulting in local distortions in contours of metal lines based on multi-modal (global-local) feature fusion.

We develop the new neural network architecture to improve execution speedaccuracy. The result shows that our AI detection system can run on a microcomputer in real-time, meanwhile its accuracy 90. We put that microcomputer on a dronereturn the result to smartphone App in real-time.

Low-power deep learning accelerator integrates neural network design / training, model compressionaccelerator design. It uses a small amount of computationmemory footprint to realize high performance on the edge device. The image semantic segmentation technology can reach 80 frames per second (resolution:1024*2048) to meet real-time requirements for autonomous driving.

Based on deep learning camera/radar object detectiontracking technology, the proposed road side unit (RSU) system has achieved over 95 vehicle detection accuracy within 100m detection range in the processing performance of 10fps under nVidia Jetson Xavier platform. Compared to the 32-beam lidar based RSU, the proposed RSU achieves 97 reduction of sensor cost that exhibits high competitiveness in deployment cost. The proposed RSU system has been verified in fieldswe are now cooperating with an industry partner to deploy the RSU system in both TainanTao-Yuan cities.