Ying Wan | Searcharxiv

arXiv2025

Hadamard-Based Recursive Aperture Decoded Ultrasound Imaging (READI) With Estimated Motion-Compensated Compounding (EMC2) Using Top-Orthogonal to Bottom Electrode (TOBE) Arrays

Hadamard matrix-based aperture encoding is a method for producing synthetic aperture datasets with high Signal-to-Noise Ratios. Recently, the pulse inversion capabilities of bias-sensitive Top-Orthogonal to Bottom Electrode (TOBE) arrays have driven the development of multiple Hadamard-based sequences. These sequences produce high-quality static images but are sensitive to motion. This work introduces Recursive Aperture Decoded Imaging (READI) and Estimated Motion-Compensated Compounding (EMC2), which look to reduce this sensitivity. READI is a novel decoding and beamforming technique for Hadamard aperture-encoded sequences that produces multiple low-resolution images from subsets of the full sequence. These READI images are less affected by motion and sum to form the complete high-resolution image. EMC2 describes the process of comparing these low-resolution images to estimate the underlying motion, then warping them to align before compounding. This produces a high-resolution image that is resiliant to motion. READI with EMC2 applied to the TOBE-based Fast Orthogonal Row-Column Electronic Scanning (FORCES) sequence. It is shown to fully restore images corrupted by probe motion and to recover tissue speckle and boundaries in images of a beating heart phantom. READI low-resolution images by themselves are demonstrated to be a marked improvement over a sparse Hadamard scheme with the same transmit count, and are able to recover blood speckle at a flow rate of 42 cm/s.

eess.IV↗

arXiv2018

DFTerNet: Towards 2-bit Dynamic Fusion Networks for Accurate Human Activity Recognition

Deep Convolutional Neural Networks (DCNNs) are currently popular in human activity recognition applications. However, in the face of modern artificial intelligence sensor-based games, many research achievements cannot be practically applied on portable devices. DCNNs are typically resource-intensive and too large to be deployed on portable devices, thus this limits the practical application of complex activity detection. In addition, since portable devices do not possess high-performance Graphic Processing Units (GPUs), there is hardly any improvement in Action Game (ACT) experience. Besides, in order to deal with multi-sensor collaboration, all previous human activity recognition models typically treated the representations from different sensor signal sources equally. However, distinct types of activities should adopt different fusion strategies. In this paper, a novel scheme is proposed. This scheme is used to train 2-bit Convolutional Neural Networks with weights and activations constrained to {-0.5,0,0.5}. It takes into account the correlation between different sensor signal sources and the activity types. This model, which we refer to as DFTerNet, aims at producing a more reliable inference and better trade-offs for practical applications. Our basic idea is to exploit quantization of weights and activations directly in pre-trained filter banks and adopt dynamic fusion strategies for different activity types. Experiments demonstrate that by using dynamic fusion strategy can exceed the baseline model performance by up to ~5% on activity recognition like OPPORTUNITY and PAMAP2 datasets. Using the quantization method proposed, we were able to achieve performances closer to that of full-precision counterpart. These results were also verified using the UniMiB-SHAR dataset. In addition, the proposed method can achieve ~9x acceleration on CPUs and ~11x memory saving.

cs.LG↗