Weakly supervised semantic segmentation (WSSS) using only image-level labels can greatly reduce the annotation cost and therefore has attracted considerable research interest. However, its performance is still inferior to fully counterparts. To mitigate gap, we propose a saliency guided self-attention network (SGAN) address WSSS problem. The introduced mechanism able capture rich extensive contextual information but may mis-spread attentions unexpected regions. In order enable this work...

Due to the high precision and non-contact characteristics, infrared thermography has been widely used in equipment inspection ensure safety of electric power systems. A fundamental step toward automatic diagnosis is detection thermal images. Therefore, this paper presents a deep learning approach detect parts real-time. Specifically, we propose convolutional neural network that predicts coordinates, orientation angle, class type each part. prior concerning consistency between also...

Tremendous progress has been made on face detection in recent years using convolutional neural networks. While many detectors use designs designated for detecting faces, we treat as a generic object task. We implement detector based the YOLOv5 and call it YOLO5Face. make few key modifications to optimize detection. These include adding five-point landmark regression head, stem block at input of backbone, smaller-size kernels SPP, P6 output PAN block. design different model sizes, from an...

In response to the insufficient research on damage weakening evolution law of surrounding rock during roadway (tunnel) excavation and retreat, as well inadequate understanding anchoring mechanism bolts in such masses, a series triaxial loading-unloading uniaxial reloading experiments were conducted sandstone specimens using RMT-150 C mechanics testing system. Microstructural characteristics damaged mass investigated through scanning electron microscope (SEM) analysis, revealing rock. Through...

Photoplethysmography (PPG) is a non-invasive physiological monitoring technology used in health monitoring, but traditional devices have limitations infant monitoring. This paper aims to improve the comfort and reliability of by combining PPG with core-sheath textile technology. Core-sheath structural fibers are as carriers for sensors, resulting comfortable heart rate Signal processing algorithms like wavelet transform, ICA, CNN denoise signals extract rate. The proposed theory outperforms...

Abstract Since the outbreak of COVID-19 corona virus in late 2019, it has become a tremendous threat to whole world. Driven by mission save lives, we develop fever screening and tracing system which can detect patients with symptom, identify using face recognition. In addition, our big data AI platform enables possible. A real-time alert sent personnel on duty web or mobile app activates action trace patient close contacts providing an effective means control spread virus.

We present the improved network architecture, data augmentation, and training strategies for Webface track Insightface/Glint360K of masked face recognition challenge ICCV2021. One key goals is to have a balanced performance standard recognition. In order prevent overfitting recognition, we control total number faces by not more than 10% in dataset. propose few changes including new stem unit, drop block, detection alignment using YOLO5Face, feature concatenation, cycle cosine learning rate...

Abstract Cross‐dataset object detection methods can adapt to the needs of rapid category expansion in tasks. However, these are prone generate dataset‐aware errors with false alarm objects. This study is aimed address issues. A box‐separated multiple‐head module and loss function based on YOLOv8 network devised achieve cross‐dataset detection. Additionally, a sameclass‐aware fusion avoid gradient conflicts due cross‐category developed. reduce number objects caused by errors. global...

Weakly supervised semantic segmentation (WSSS) using only image-level labels can greatly reduce the annotation cost and therefore has attracted considerable research interest. However, its performance is still inferior to fully counterparts. To mitigate gap, we propose a saliency guided self-attention network (SGAN) address WSSS problem. The introduced mechanism able capture rich extensive contextual information but may mis-spread attentions unexpected regions. In order enable this work...

Software faults constantly appear during software development and evolution. The information on various platforms for bug knowledge recording, such as Stack Overflow, is mostly stored in weak entity relational database missing linkable relationships, which results negative impacts reuse. To enrich the relationships between entities construct a fault graph, we improve SALKU model by considering direction of prediction pair units, utilize it to predict class units. Experiment show improved...

We present the improved network architecture, data augmentation, and training strategies for Webface track Insightface/Glint360K of masked face recognition challenge ICCV2021. One key goals is to have a balanced performance standard recognition. In order prevent overfitting recognition, we control total number faces by not more than 10\% in dataset. propose few changes including new stem unit, drop block, detection alignment using YOLO5Face, feature concatenation, cycle cosine learning rate,...