Real-world environmental conditions such as low-light, haze, and underwater typically result in degraded images characterized by low contrast, color distortion, inconsistent lighting, and high noise levels. This paper introduces a novel approach, the Double Opponency-based Transformer network (DOT), to address common limitations like local details loss and spatial misalignment in existing image enhancement methods. Moreover, unlike the traditional encoder–decoder architectures based models, through decomposing input images into different color components and employing learnable double-opponent processing and non-bottleneck Transformer network, DOT captures intricate color relationships and enhances visual quality while maintaining reasonable global brightness and color accuracy. In the experiments, we demonstrate the effectiveness of the DOT model through extensive quantitative and qualitative evaluations against state-of-the-art methods on various datasets, including low-light, underwater, and hazy image enhancement.

Transformers have achieved great success in natural language processing and computer vision. The core and basic technique of transformers is the self-attention mechanism. The vanilla self-attention mechanism has quadratic complexity, which limits its applications to vision tasks. Most of the existing linear self-attention mechanisms will sacrifice performance to some extent to reduce complexity. In this paper, we propose a novel linear approximation of the vanilla self-attention mechanism named CURSA to achieve both high performance and low complexity at the same time. CURSA is based on the CUR decomposition to decompose the multiplication of large matrices into the multiplication of several small matrices to achieve almost linear complexity. Experiment results of CURSA in image classification tasks, semantic segmentation tasks, object detection tasks, and long-range arena show that it outperforms state-of-the-art self-attention mechanisms with better data efficiency, faster speed, and higher accuracy.

Superpixels can preserve the structure and reduce the redundancy of the original image. Because of these advantages, superpixel generation or superpixel segmentation is widely used as a pre-processing step in many image processing tasks. Although superpixels can be employed to reduce computational complexity, some challenges, such as the non-Euclidean feature learning problem introduced by superpixels, still exist. This survey provides a comprehensive overview of the state-of-the-art superpixel methods, major challenges, commonly used evaluation metrics, applications of superpixels, and potential future directions for the study of superpixels. We first give a review of the state-of-the-art superpixel methods. Next, we use different evaluation metrics to evaluate the performance of 24 up-to-date superpixel methods on different datasets in different noisy environments. After that, we introduce several up-to-date applications of superpixels. Finally, we give several possible future directions for addressing the challenges of superpixels.

Yi Lu, Jiawang Cao, Yongliang Wu, Bozheng Li, Licheng Tang, Yangguang Ji, Chong Wu, Jay Wu, Wenbo Zhu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The commercial success of generative artificial intelligence (GenAI) has driven an exponential surge in demand for real-time inference in Vision Transformer (ViT) applications, including latency-sensitive domains in autonomous driving, medical imaging and computational photography. This paper introduces FastViT, a high-performance and energy-efficient hardware accelerator for emerging kernel function-based linear attention mechanisms. By leveraging cost-efficient multiplication, mixed-precision quantisation and optimised data flow, FastViT improves real-time performance for high-resolution dense prediction tasks. Compared to existing approaches, experiments demonstrate that FastViT achieves higher throughput and energy efficiency while maintaining negligible accuracy degradation and balanced resource allocation. In the future, we will improve its scalability for next-generation hardware equipped with advanced DSP cores.

The attention mechanism is the key to the success of transformers in different machine learning tasks. However, the quadratic complexity with respect to the sequence length of the vanilla softmax-based attention mechanism becomes the major bottleneck for the application of long sequence tasks, such as vision tasks. Although various efficient linear attention mechanisms have been proposed, they need to sacrifice performance to achieve high efficiency. What's more, memory-efficient methods, such as FlashAttention-1-3, still have quadratic computation complexity which can be further improved. In this paper, we propose a novel efficient linear fast attention (ELFATT) mechanism to achieve low memory input/output operations, linear computational complexity, and high performance at the same time. ELFATT offers 4-7x speedups over the vanilla softmax-based attention mechanism in high-resolution vision tasks without losing performance. ELFATT is FlashAttention friendly. Using FlashAttention-2 acceleration, ELFATT still offers 2-3x speedups over the vanilla softmax-based attention mechanism on high-resolution vision tasks without losing performance. Even in some non-vision tasks of long-range arena, ELFATT still achieves leading performance and offers 1.2-2.3x speedups over FlashAttention-2. Even on edge GPUs, ELFATT still offers 1.6x to 2.0x speedups compared to state-of-the-art attention mechanisms in various power modes from 5W to 60W. Furthermore, ELFATT can be used to enhance and accelerate diffusion tasks directly without training.

This paper proposes the Dual-Stage Sparse Attention (DuSA) mechanism for attention acceleration of transformers. In the first stage, DuSA performs intrablock sparse attention to aggregate local inductive biases. In the second stage, DuSA performs interblock sparse attention to obtain long-range dependencies. Both stages have low computational complexity and can be further accelerated by memory acceleration attention mechanisms directly, which makes DuSA faster than some extremely fast attention mechanisms. The dual-stage sparse attention design provides a lower error in approximating vanilla scaled-dot product attention than the basic single-stage sparse attention mechanisms and further advances the basic sparse attention mechanisms to match or even outperform vanilla scaled-dot product attention. Even in some plug and play situations, DuSA can still maintain low performance loss. DuSA can be used in both training and inference acceleration. DuSA achieves leading performance in different benchmarks: long range arena, image classification, semantic segmentation, object detection, text to video generation, and long context understanding, and accelerates models of different sizes.

Convolutional neural networks (CNNs) are usually used as a backbone to design methods in biomedical image segmentation. However, the limitation of receptive field and large number of parameters limit the performance of these methods. In this paper, we propose a graph neural network (GNN) based method named GNN-SEG for the segmentation of brain tissues. Different to conventional CNN based methods, GNN-SEG takes superpixels as basic processing units and uses GNNs to learn the structure of brain tissues. Besides, inspired by the interaction mechanism in biological vision systems, we propose two kinds of interaction modules for feature enhancement and integration. In the experiments, we compared GNN-SEG with state-of-the-art CNN based methods on four datasets of brain magnetic resonance images. The experimental results show the superiority of GNN-SEG.

Identification of essential proteins is a hot topic in bioinformatics. In recent years, various traditional methods have been proposed, which usually take topological features to rank the proteins and then set a threshold for selecting essential proteins. Some researchers have also tried to take machine learning methods or deep learning methods for the prediction of essential proteins. However, these methods can not well extract the topological features of protein-protein interaction (PPI) network. Besides, although some scholars proposed to combine biological information with PPI network to reduce the noise of PPI data, how to well combine the biological information with PPI network remains to be a problem. In this paper, we propose a non-local graph neural network to tackle the above problems. In our algorithm, we use graph convolutional network (GCN) layers to extract graph embeddings of PPI network and take three kinds of biological information like gene expression profiles, subcellular localization data, and protein complex data as the features of proteins. Besides, based on the characteristics of our model we design a non-local module as the attention mechanism, which is widely used in convolutional neural networks, to aggregate the information of the graph embeddings. Finally, we use another GCN layer for the node classification and obtain the essential proteins. In the experiments, we evaluated our method based on the widely used S. cerevisiae (Yeast) dataset. The experimental results show that our method can obtain significant improvements over traditional topology-based methods, machine learning-based methods, and recently proposed deep learning-based methods.

Most superpixel methods are sensitive to noise and cannot control the superpixel number precisely. To solve these problems, in this article, we propose a robust superpixel method called fuzzy simple linear iterative clustering (Fuzzy SLIC), which adopts a local spatial fuzzy C-means clustering and dynamic fuzzy superpixels. We develop a fast and precise superpixel number control algorithm called onion peeling (OP) algorithm. Fuzzy SLIC is insensitive to most types of noise, including Gaussian, salt and pepper, and multiplicative noise. The OP algorithm can control the superpixel number accurately without reducing much computational efficiency. In the validation experiments, we tested the Fuzzy SLIC and OP algorithm and compared them with state-of-the-art methods on the BSD500 and Pascal VOC2007 benchmarks. The experiment results show that our methods outperform state-of-the-art techniques in both noise-free and noisy environments.

Using graph theory to identify essential proteins is a hot topic at present. These methods are called network-based methods. However, the generalization ability of most network-based methods is not satisfactory. Hence, in this paper, we consider the identification of essential proteins as a multi-objective optimization problem and use a novel multi-objective optimization method to solve it. The optimization result is a set of Pareto solutions. Every solution in this set is a vector which has a certain number of essential protein candidates and is considered as an independent predictor or voter. We use a voting strategy to assemble the results of these predictors. To validate our method, we apply it on the protein-protein interactions (PPI) datasets of two species (Yeast and Escherichia coli). The experiment results show that our method outperforms state-of-the-art methods in terms of sensitive, specificity, F-measure, accuracy, and generalization ability.

Superpixel is widely used in image processing. And among the methods for superpixel generation, clustering-based methods have a high speed and a good performance at the same time. However, most clustering-based superpixel methods are sensitive to noise. To solve these problems, in this paper, we first analyze the features of noise. Then according to the statistical features of noise, we propose a novel centroid update approach to enhance the robustness of clustering-based superpixel methods. Besides, we propose a novel superpixel-based edge detection method. The experiments on BSD500 dataset show that our approach can significantly enhance the performance of clustering-based superpixel methods in noisy environment. Moreover, we also show that our proposed edge detection method outperforms other classical methods.

Superpixel-based fast fuzzy C-means clustering (SFFCM) is an efficient method for color image segmentation. However, it is sensitive to noise and blur. Its superpixel method called multiscale morphological gradient reconstruction (MMGR) is time consuming. In this paper, we propose an improved SFFCM method (ISFFCM) which replaces the MMGR in SFFCM with fuzzy simple linear iterative clustering (Fuzzy SLIC). Fuzzy SLIC is faster and more robust than MMGR for most types of noise, including salt and pepper noise, Gaussian noise and multiplicative noise. It is also more robust to image blur. In the validation experiments, we tested ISFFCM and SFFCM on the Berkeley benchmark. The experiment results show that our method outperforms SFFCM under noise and blurring environments.