YOLOSR-IST: A deep learning method for small target detection in infrared remote sensing images based on super-resolution and YOLO

Signal Processing, Journal Year: 2023, Volume and Issue: 208, P. 108962 - 108962

Published: Feb. 5, 2023

Language: Английский

---

Continuous transfer of neural network representational similarity for incremental learning

Neurocomputing, Journal Year: 2023, Volume and Issue: 545, P. 126300 - 126300

Published: May 13, 2023

Language: Английский

---

Efficient Lung Cancer Image Classification and Segmentation Algorithm Based on an Improved Swin Transformer

Electronics, Journal Year: 2023, Volume and Issue: 12(4), P. 1024 - 1024

Published: Feb. 18, 2023

With the advancement of computer technology, transformer models have been applied to field vision (CV) after their success in natural language processing (NLP). In today’s rapidly evolving medical field, radiologists continue face multiple challenges, such as increased workload and diagnostic demands. The accuracy traditional lung cancer detection methods still needs be improved, especially realistic scenarios. this study, we evaluated performance Swin Transformer model classification segmentation cancer. results showed that pre-trained Swin-B achieved a top-1 82.26% mission, outperforming ViT by 2.529%. Swin-S demonstrated improvement over other terms mean Intersection Union (mIoU). These suggest pre-training can an effective approach for improving these tasks.

Language: Английский

---

ICGNet: An intensity-controllable generation network based on covering learning for face attribute synthesis

Information Sciences, Journal Year: 2024, Volume and Issue: 660, P. 120130 - 120130

Published: Jan. 21, 2024

Language: Английский

---

High Speed and Accuracy of Animation 3D Pose Recognition Based on an Improved Deep Convolution Neural Network

Applied Sciences, Journal Year: 2023, Volume and Issue: 13(13), P. 7566 - 7566

Published: June 27, 2023

Pose recognition in character animations is an important avenue of research computer graphics. However, the current use traditional artificial intelligence algorithms to recognize animation gestures faces hurdles such as low accuracy and speed. Therefore, overcome above problems, this paper proposes a real-time 3D pose system, which includes both facial body poses, based on deep convolutional neural networks further designs single-purpose estimation system. First, we transformed human extracted from input image abstract data structure. Subsequently, generated required at runtime dataset. This challenges conventional concept monocular estimation, extremely difficult achieve. It can also achieve running speed resolution 384 fps. The proposed method was used identify multiple-character using multiple datasets (Microsoft COCO 2014, CMU Panoptic, Human3.6M, JTA). results indicated that improved algorithm performance by approximately 3.5% 8–10 times, respectively, significantly superior other classic algorithms. Furthermore, tested system pose-recognition datasets. attitude reach 24 fps with error 100 mm, considerably less than 2D 60 learning study yielded surprisingly performance, proving deep-learning technology for has great potential.

Language: Английский

---

A new hybrid image denoising algorithm using adaptive and modified decision-based filters for enhanced image quality

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: March 15, 2025

Denoising is one of the most important processes in digital image processing to recover visual quality and structural integrity images. Traditional methods often suffer from limitations like computational complexity, over-smoothing, inability preserve critical details, particularly edges. This paper introduces a hybrid denoising algorithm combining Adaptive Median Filter (AMF) Modified Decision-Based (MDBMF) address these challenges. The AMF adjusts window sizes dynamically precisely detect noisy pixels, MDBMF selectively recovers corrupted pixels without affecting intact regions, effectively reducing noise while preserving subjective analysis supplemented with objective analyses which proves that approach performance considerably outperforms existing state-of-the-art methods. test conducted on nine benchmark images standard medical dataset, namely, Chest Liver different densities range 10 90%. Quantitative evaluations PSNR, MSE, IEF, SSIM, FOM VIF clearly show superiority when compared approaches. improvement PSNR was up 2.34 dB, IEF more than 20%, MSE 15% over other BPDF, AT2FF, SVMMF. Improvement values SSIM 0.07, confirms improved similarity. Furthermore, metrics demonstrate remarkable approach: both exceeded all techniques evaluated, reaching 0.68 0.61, respectively.

Language: Английский

---

A Financial Time-Series Prediction Model Based on Multiplex Attention and Linear Transformer Structure

Applied Sciences, Journal Year: 2023, Volume and Issue: 13(8), P. 5175 - 5175

Published: April 21, 2023

Financial time-series prediction has been an important topic in deep learning, and the of financial time series is great importance to investors, commercial banks regulators. This paper proposes a model based on multiplexed attention mechanisms linear transformers predict series. The transformer faster training efficiency long-time forecasting capability. Using reduces original transformer’s complexity preserves decoder’s mechanism. results show that proposed method can effectively improve accuracy model, increase inference speed reduce number operations, which new implications for

Language: Английский

---

3D human pose and shape estimation via de-occlusion multi-task learning

Neurocomputing, Journal Year: 2023, Volume and Issue: 548, P. 126284 - 126284

Published: May 3, 2023

Language: Английский

---

Wind power prediction based on WT-BiGRU-attention-TCN model

Frontiers in Energy Research, Journal Year: 2023, Volume and Issue: 11

Published: April 13, 2023

Accurate wind power prediction is crucial for the safe and stable operation of grid. However, generation has large random volatility intermittency, which increases difficulty prediction. In order to construct an effective model based on achieve grid dispatch after connected grid, a WT-BiGRU-Attention-TCN proposed. First, wavelet transform (WT) used reduce noises sample data. Then, temporal attention mechanism incorporated into bi-directional gated recurrent unit (BiGRU) highlight impact key time steps results while fully extracting features context. Finally, performance enhanced by further more high-level through convolutional neural network (TCN). The show that our proposed outperforms other baseline models, achieving root mean square error 0.066 MW, absolute percentage 18.876%, coefficient determination (R 2 ) reaches 0.976. It indicates noise-reduction WT technique can significantly improve performance, also shows using TCN accuracy.

Language: Английский

---

Real-time load forecasting model for the smart grid using bayesian optimized CNN-BiLSTM

Frontiers in Energy Research, Journal Year: 2023, Volume and Issue: 11

Published: May 5, 2023

A smart grid is a new type of power system based on modern information technology, which utilises advanced communication, computing and control technologies employs sensors, measurement, communication devices that can monitor the status operation various in real-time optimise dispatch through intelligent algorithms to achieve efficient system. However, due its complexity uncertainty, how effectively perform prediction an important challenge. This paper proposes model attention mechanism convolutional neural network (CNN) combined with bi-directional long short-term memory BiLSTM.The has stronger spatiotemporal feature extraction capability, more accurate capability better adaptability than ARMA decision trees. The traditional models tree often only use simple statistical methods for prediction, cannot meet requirements high accuracy efficiency load so CNN-BiLSTM Bayesian optimisation following advantages suitable compared tree. CNN hierarchical structure containing several layers such as layer, pooling layer fully connected layer. mainly used extracting features from data images, dimensionality reduction features, classification recognition. core operation, locally weighted summation input extract data. In convolution different be extracted by setting kernels BiLSTM capture semantic dependencies both directions. consists two LSTM process sequence forward backward directions combine obtain comprehensive contextual information. access front back inputs at each time step results. It prevents gradient explosion disappearance while capturing longer-distance dependencies. extracts then optimises them Bayes. By collecting system, including power, load, weather other factors, our uses deeply learn grids key future prediction. Meanwhile, algorithm model’s hyperparameters, thus improving performance. provide reference help energy utilisation

Language: Английский

---