A5103031514- Advanced Neural Network Applications
- CCD and CMOS Imaging Sensors
- Advanced Memory and Neural Computing
- COVID-19 diagnosis using AI
- EEG and Brain-Computer Interfaces
- Digital Holography and Microscopy
- Optical measurement and interference techniques
- Advanced Image and Video Retrieval Techniques
- Advanced X-ray Imaging Techniques
- Machine Learning in Materials Science
- Semiconductor Lasers and Optical Devices
- Radiomics and Machine Learning in Medical Imaging
- Advanced Chemical Sensor Technologies
- Photonic and Optical Devices
- Currency Recognition and Detection
- Neuroscience and Neural Engineering
- Semiconductor materials and devices
- Video Surveillance and Tracking Methods
- Context-Aware Activity Recognition Systems
- Visual Attention and Saliency Detection
- Water Quality Monitoring Technologies
Helmholtz-Zentrum Hereon
2024
ETH Zurich
2023-2024
Università della Svizzera italiana
2023
University of Bologna
2023
This paper introduces a highly flexible, quantized, memory-efficient, and ultra-lightweight object detection network, called TinyissimoYOLO. It aims to enable on microcontrollers in the power domain of milliwatts, with less than 0.5 MB memory available for storing convolutional neural network (CNN) weights. The proposed quantized architecture 422 k parameters, enables real-time embedded microcontrollers, it has been evaluated exploit CNN accelerators. In particular, deployed MAX78000...
This paper deploys and explores variants of TinyissimoYOLO, a highly flexible fully quantized ultra-lightweight object detection network designed for edge systems with power envelope few milliwatts. With experimental measurements, we present comprehensive characterization the network's performance, exploring impact various parameters, including input resolution, number classes, hidden layer adjustments. We deploy TinyissimoYOLO on state-of-the-art ultra-low-power extreme platforms,...
Advances in lightweight neural networks have revolutionized computer vision a broad range of IoT applications, encompassing remote monitoring and process automation. However, the detection small objects, which is crucial for many these remains an underexplored area current research, particularly low-power embedded devices that host resource-constrained processors. To address said gap, this paper proposes adaptive tiling method energy-efficient object networks, including YOLO-based models...
This study presents a deep learning algorithm for solving the inverse problem of phase retrieval in optical near-field (Fresnel) regime using single intensity measurement (in-line hologram). The was developed self-learning fashion based on generative adversarial networks (GANs) tomographic reconstruction (GANrec). In this paper, original GANrec is adapted to solve problem. From hologram, can recover both and amplitude unpropagated exit wave field. Compared with other state-of-the-art...
Smart glasses are rapidly gaining advanced functionality thanks to cutting-edge computing technologies, accelerated hardware architectures, and tiny AI algorithms. Integrating into smart featuring a small form factor limited battery capacity is still challenging when targeting full-day usage for satisfactory user experience. This paper illustrates the design implementation of machine-learning algorithms exploiting novel low-power processors enable prolonged continuous operation in glasses....
This paper deploys and explores variants of TinyissimoYOLO, a highly flexible fully quantized ultra-lightweight object detection network designed for edge systems with power envelope few milliwatts. With experimental measurements, we present comprehensive characterization the network's performance, exploring impact various parameters, including input resolution, number classes, hidden layer adjustments. We deploy TinyissimoYOLO on state-of-the-art ultra-low-power extreme platforms,...
Human Activity Recognition (HAR) is an attractive topic to perceive human behavior and supplying assistive services. Besides the classical inertial unit vision-based HAR methods, new sensing technologies, such as ultrasound body-area electric fields, have emerged in enhance user experience accommodate application scenarios. As those sensors are often paired with AI for HAR, they frequently encounter challenges due limited training data compared more widely IMU or solutions. Additionally,...
This study presents a deep learning algorithm for solving the inverse problem of phase retrieval in optical near-field (Fresnel) regime using single intensity measurement (in-line hologram). The was developed self-learning fashion based on generative adversarial networks (GANs) tomographic reconstruction (GANrec). In this paper, original GANrec is adapted to solve problem. From hologram, can recover both and amplitude unpropagated exit wave field. Compared with other state-of-the-art...
Electroencephalogram (EEG)-based Brain-Computer Interfaces (BCIs) have garnered significant interest across various domains, including rehabilitation and robotics. Despite advancements in neural network-based EEG decoding, maintaining performance diverse user populations remains challenging due to feature distribution drift. This paper presents an effective approach address this challenge by implementing a lightweight efficient on-device learning engine for wearable motor imagery...
Currently, most bone implants used in orthopedics and traumatology are non-degradable may need to be surgically removed later on e.g. the case of children. This removal is associated with health risks which could minimized by using biodegradable implants. Therefore, research magnesium-based ongoing, can objectively quantified through synchrotron radiation microtomography subsequent image analysis. In order evaluate suitability these materials, their stability over time, accurate pixelwise...
This paper introduces a highly flexible, quantized, memory-efficient, and ultra-lightweight object detection network, called TinyissimoYOLO. It aims to enable on microcontrollers in the power domain of milliwatts, with less than 0.5MB memory available for storing convolutional neural network (CNN) weights. The proposed quantized architecture 422k parameters, enables real-time embedded microcontrollers, it has been evaluated exploit CNN accelerators. In particular, deployed MAX78000...
Event-based cameras, also called silicon retinas, potentially revolutionize computer vision by detecting and reporting significant changes in intensity asynchronous events, offering extended dynamic range, low latency, power consumption, enabling a wide range of applications from autonomous driving to longtime surveillance. As an emerging technology, there is notable scarcity publicly available datasets for event-based systems that feature frame-based order exploit the benefits both...
Event-based cameras, also called silicon retinas, potentially revolutionize computer vision by detecting and reporting significant changes in intensity asynchronous events, offering extended dynamic range, low latency, power consumption, enabling a wide range of applications from autonomous driving to longtime surveillance. As an emerging technology, there is notable scarcity publicly available datasets for event-based systems that feature frame-based order exploit the benefits both...
This paper addresses the growing interest in deploying deep learning models directly in-sensor. We present "Q-Segment", a quantized real-time segmentation algorithm, and conduct comprehensive evaluation on low-power edge vision platform with an in-sensors processor, Sony IMX500. One of main goals model is to achieve end-to-end image for vessel-based medical diagnosis. Deployed IMX500 platform, Q-Segment achieves ultra-low inference time in-sensor only 0.23 ms power consumption 72mW. compare...