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.

"We will demonstrate the following three technical achievements of our joint project: 1. Deployment of HarDNet on GPU (power consumption: 200 Watts) 2. Deployment of HarDNet on FPGA (power consumption: several tens of Watts) [winning 2nd place in the FPGA track, LPCVC 2020] 3. Deployment of HarDNet on lightweight edge devices such as Raspberry Pi (power consumption: single-digit, 10 Watts) [winning 3rd place in the DSP track4th place in the CPU track, LPCVC 2020]"

1. CNN model architecture optimization techniques: Propose an Agilev3-L architecture, with a performance index of mAP@50:95FPS:FP16 up to 171.79 2. GOP-mode acceleration scheme for real time inference: employ GOP-modeTRACKING algorithms to enhance the processing rate. 3. Propose a rapid data labeling system HiTag.

This system help doctors to distinguish age-related macular degeneration (ARMD)polypoidal choroidal vasculopathy (PCV) by color fundus photography. We have collected large image dataset from Taipei Veterans General Hospitalother hospitals to establishvalidate CNN model, which had been tested by external image datasetmodified to improve the accuracy of the AI model.

Use NLP techniques to realize the automatic coding of ICD10. According to the input of the patient’s age, gender, medical order, admissions, progress note, surgical records, discharge, ICD-10 diagnostic codeICD-10 disposal code, perform machine learning model trainingcode prediction. In addition, the combination codemedical order-related coding rules in practice are used to establish corresponding rules to optimize the accuracy of AI prediction.

Similar to the critical point tracking technology of human vision, the split attentionspatial grouping enhancement module, combining the multi-path grouping architecturespatial attention technology, can accurately extract important information from the imageimprove the network performance. Moreover, adopting neural architecture search technology to automatically search for the most suitable network architecture based on current moduleshardware devices can balance diagnosis speedhigh accuracy.

"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."

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.

The technology aims to develop an AI-based integrated system for emotional detections (anger, sadness, happiness,neutral)multimodal physiological signals (ECG, EEG,PPG),apply to patients with cardiovascular disease. Patients monitor their emotionalphysical statusadminister bio-neuro-feedback to improve their well-being, track disease progression,prevent adverse prognosis.

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.

Integrating deep learning, 3D information analysis, hyperspectral analysis, computer vision analysiscombined the multi-source images to develop fruit quality monitoring techniques, including: planting area monitoring, position monitoring, yield monitoring, harvesting time prediction, fruit maturity measurement, fruit quality testing, to achieve the goal of enhancing industrial value.

This project integrated with industrial foresight technologies, including UAV, artificial intelligenceimage recognition, to collect real-time images, apply algorithm in evaluation, link the technology of IOT (Internet of Things)environment sensing,use unmanned vehicle to conduct controlling work.

Constructing a functional connectome and its computational model is a crucial step toward understanding the mechanisms of brain functions. To achieve this goal, we developed two correlated technologies: (1) An all-optical physiology (AOP) that is capable of millisecond volumetric imaging and accurate stimulation in living animal brains. This system allows us to establish functional connectome and neural coding with a single-cell resolution. (2) A cellular-level spiking neural circuit simulation system that is capable of tuning itself based on the input data from the AOP system. We have demonstrated our technologies in the Drosophila late visual system and will apply them in the brains of larger species such as mice. Our technologies will greatly enhance knowledge of brain operation.

We propose a semantic-guided framework (FashionOn+) that generates image-based virtual try-on results with arbitrary poses. FashionOn+ contains three stages: (I) conducts the semantic segmentation to have the prior knowledge of body parts for rendering the corresponding texture in stage (II). (III) refines two salient regions, i.e., faceclothes, to generate high-quality results. With the novel architecture, we win first place in the Multi-pose Virtual Try-on Challenge in CVPR, 2020. Further, we tackle the low-resolution limitation (256x192)achieve high-resolution results (640x480).

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.

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

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.

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.

The computer-aided nodule detection system in CT image consists of the search sliding window, YOLOv2 architecture, 3-D CapsNet, skip connection,post-processing. First, the CT image is divided into numerous VOIs by sliding window. Second, a 3-D CapsNet based on YOLOv2 architectureskip connection is applied to the VOIs for classifying VOIs as nodulenot. Finally, the non-maximum suppression algorithm is performed to decide the final detection result.

We have developed a movement assistive system using EEG-controlled functional electric stimulation system. Users’ movement intentions are recognized by brain computer interface, implemented by deep learning network,the detected users’ intentions are then translated into commands to trigger a functional electric stimulation (FES) device to activate users’ particular movements, such as grasping, hand raising, holding a cup, etc. In our study, we have designed our own wireless dry-electrode EEG systemour own FES system.

Artificial intelligence 3D sensing image processing system based on array sensing Lidar aims to construct 3D image with high-quality immersion for AR/VR. The developed 3D scene recording system is based on the color camerahigh-accuracy chaotic LiDAR. The chaotic Lidar with APD arrayTOF sensors supports millimeter-accuracyinterference-avoiding capability within 100 meter in both indooroutdoor environments. High-performance embedded CNN processor supports high-throughput, high energy-efficient,low DRAM bandwidth computations for various image AI applications.

PANCREASaver contains a “PC automatic segmentation model” (image segmentation)a “PC analysis AI model” (image classification) that can read the DICOM format of postcontrast CT images directly for the automatic analysis process. After conducting prep-processing with image processing algorithms, C2FNAS is employed to illustrate the tumor position prior to the diagnosis conducted by CNN. The results can be provided to the physician for diagnostic reference so as to reduce early omissionsincrease the detection rate of pancreatic cancer.

High-accuracy 3D sensingAI image processing system for constructing high-quality immersion 3D image for AR/VR. Chaotic Lidar with APD arrayTOF sensors supports millimeter-accuracyinterference-avoiding capability. High-performance CNN processor supports high-performancelow DRAM bandwidth computations for various image AI applications.

1.Low-Power Processing-in-sensor CMOS Image Sensor 2.Implementation of Low-PowerLow-Latency Deep Learning Chip based on Neuromorphic Intelligence (collaborated with MOST 108-2622-8-007-009) 3.Development of Neuromorphic Chip based on The Fruit Fly VisualSpatial Sensory Systems 4.Hardware-Software Co-Design for Neuromorphic AI Chips

Assisted by in-house AI deep compression toolchain (ezLabel, ezModel, ezQUANT, ezHybrid-M), the proposed technology supports automatic AI model designoptimization with the integrated performance of 120x model size reduction70x power reduction in 2D CNN model,develops a world-first 1/2/4/8-bit CNN model realized by the developed high efficiency Hybrid fixed point CNN NPU (Hybrid-NPU), which has been verified in Xilinx ZCU102 FPGAachieves the performance up to 2.5 TOPS(8-b)/ 20TOPS(1-b)@28nm technology running at 550MHz4TOPS/W energy efficiency.