Transformer-style convolution network for lightweight image super-resolution

Multimedia Tools and Applications, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 4, 2025

Language: Английский

---

Double Attention: An Optimization Method for the Self-Attention Mechanism Based on Human Attention

Biomimetics, Journal Year: 2025, Volume and Issue: 10(1), P. 34 - 34

Published: Jan. 8, 2025

Artificial intelligence, with its remarkable adaptability, has gradually integrated into daily life. The emergence of the self-attention mechanism propelled Transformer architecture diverse fields, including a role as an efficient and precise diagnostic predictive tool in medicine. To enhance accuracy, we propose Double-Attention (DA) method, which improves neural network's biomimetic performance human attention. By incorporating matrices generated from shifted images mechanism, network gains ability to preemptively acquire information surrounding regions. Experimental results demonstrate superior our approaches across various benchmark datasets, validating their effectiveness. Furthermore, method was applied patient kidney datasets collected hospitals for diabetes diagnosis, where they achieved high accuracy significantly reduced computational demands. This advancement showcases potential methods field biomimetics, aligning well goals developing innovative bioinspired tools.

Language: Английский

---

A Pyramid Fusion MLP for Dense Prediction

IEEE Transactions on Image Processing, Journal Year: 2025, Volume and Issue: 34, P. 455 - 467

Published: Jan. 1, 2025

Recently, MLP-based architectures have achieved competitive performance with convolutional neural networks (CNNs) and vision transformers (ViTs) across various tasks. However, most methods introduce local feature interactions to facilitate direct adaptation downstream tasks, thereby lacking the ability capture global visual dependencies multi-scale context, ultimately resulting in unsatisfactory on dense prediction. This paper proposes a effective architecture called Pyramid Fusion MLP (PFMLP) address above limitation. Specifically, each block PFMLP introduces pooling fully connected layers generate pyramids, which are subsequently fused using up-sample an additional layer. Employing different down-sample rates allows us obtain diverse receptive fields, enabling model simultaneously long-range fine-grained cues, exploiting potential of context information enhancing spatial representation power model. Our is first lightweight comparable results state-of-the-art CNNs ViTs ImageNet-1K benchmark.With larger FLOPs, it exceeds CNNs, ViTs, MLPs under similar computational complexity. Furthermore, experiments object detection, instance segmentation, semantic segmentation demonstrate that acquired from can be seamlessly transferred producing results. All materials contain training codes logs released at https://github.com/huangqiuyu/PFMLP.

Language: Английский

---

Activity-aware electrocardiogram biometric verification utilising deep learning on wearable devices

EURASIP Journal on Information Security, Journal Year: 2025, Volume and Issue: 2025(1)

Published: Feb. 25, 2025

Language: Английский

---

Reverse design of load-bearing broadband metamaterial absorber assisted by deep learning

Smart Materials and Structures, Journal Year: 2024, Volume and Issue: 33(12), P. 125029 - 125029

Published: Nov. 17, 2024

Abstract In response to the current challenges of narrow absorption bandwidth, weak load-bearing capacity, and low design efficiency in absorbing structures, this study focuses on reverse broadband metamaterial absorber. A parameterized model absorber was developed by integrating composite sandwich structure with electromagnetic resonant layers. The layer constructed using combination Vicsek-fractal circular rings, resistive films employed broaden bandwidth. deep learning-based forward prediction established accurately predict absorbance shapley additive explanations (SHAP) framework utilized analyze network, revealing influence various parameters at center frequencies across L K band spectrum. Additionally, group teaching optimization algorithm (GTOA) introduced into process, leading development an automated method for that can achieve specific objectives. Using GTOA-based method, a capable effectively vertically incident waves within 3–20 GHz frequency range designed. designed fabricated, its performance measured arch method. measurement results were found be good agreement simulation data. mechanism analyzed based calculation equivalent resonance observed frequency. It determined primary effect is induced electric triggered waves. proposed applied radar stealth military targets such as naval vessels. research methodology approach demonstrate significant generalizability engineering applicability.

Language: Английский

---

Semantic image representation for image recognition and retrieval using multilayer variational auto-encoder, InceptionNet and low-level image features

The Journal of Supercomputing, Journal Year: 2024, Volume and Issue: 81(1)

Published: Dec. 27, 2024

Language: Английский

---

MCM-ViT:Mask-guided context-enhanced multi-scale transformer for fine-grained visual classification

Computers & Electrical Engineering, Journal Year: 2024, Volume and Issue: 122, P. 109888 - 109888

Published: Dec. 5, 2024

Language: Английский

---