The key to improving the performance of graph convolutional networks (GCN) is to fully explore the correlation between neighboring and distant information. Aiming at the over-smoothing problem of GCN, in order to make full use of the relationship among features, graphs and labels, a graph residual generation network based on multi-information aggregation (MIA-GRGN) is proposed. Firstly, aiming at the defects of GCN, we design a deep initial residual graph convolution network (DIRGCN), which connects the initial input through residuals, so that each layer node retains part of the information of the initial features, ensuring the localization of the graph structure and effectively alleviating the problem of over-smoothing. Secondly, we propose a random graph generation method (RGGM) by utilizing graph edge sampling and negative edge sampling, and optimize the supervision loss function of DIRGCN in the form of generation framework. Finally, applying RGGM and DIRGCN as inference modules for modeling hypotheses and obtaining approximate posterior distributions of unknown labels, an optimized loss function is obtained, we construct a multi-information aggregation MIA-GRGN that combines graph structure, node characteristics and label joint distribution. Experiments on benchmark graph classification datasets show that MIA-GRGN achieves better classification results compared with the benchmark models and mainstream models, especially for datasets with less dense edge relationships between nodes.

Mining sentiment from user-generated social media content is crucial for fields such as business decision-making and public opinion analysis. Aspect-based sentiment analysis (ABSA) aims to fine-grainedly identify the sentiment of different entities or attributes in texts. However, existing methods face two challenges: 1) attention mechanisms are prone to interference from aspect-irrelevant words, leading to attention bias; and 2) irrelevant syntactic dependencies are frequently introduced by graph convolutional networks constructed upon dense dependency trees, causing semantic interference. To address these challenges, this article proposes a novel method that combines graph transformer with aspect-oriented graph convolutional network (GT-AGCN), reducing attention bias and semantic interference while modeling global semantics and local syntax. First, a context encoder is constructed to capture contextualized semantic representations of the input text sequences. Second, to efficiently model global semantic information, a syntactically constrained Transformer module is developed to alleviate the inherent attention bias. Meanwhile, to accurately capture local syntax information, an aspect-oriented graph convolutional network is designed to suppress interference from aspect-irrelevant terms. Finally, a classifier is utilized to predict the sentiment polarities of target aspects by combining global semantic features with local syntactic features. Empirical evaluations across five widely-adopted ABSA benchmarks demonstrate that GT-AGCN outperforms state-of-the-art baselines, validating the effectiveness of model.

Surface defects in steel exhibit diverse morphologies and significant size variations, often accompanied by low contrast and complex background interference, presenting substantial challenges for detection tasks. To address this, we propose a cross-domain modulation-based end-to-end detection network, CDCM-DETR, to achieve efficient and precise identification of complex defects. First, we design the echo refinement guided network as the backbone, it employs an adaptive joint screening module to perceive defect regions while mitigating deep information decay. Then, we construct the Cross-Domain Modulation and Fusion architecture. It enhances high-frequency responses through frequency-domain compensation using the spectrum resampling alignment unit and strengthens the hierarchical expression of features in the spatial domain by building multi-scale pathways with the multi-kernel perception mapping module. This synergistic approach achieves consistent enhancement of cross-domain features. Experimental results on the public datasets NEU-DET and GC10-DET demonstrate that CDCM-DETR achieves the mAP50 improvement of 4.6% and 4.3%, respectively, over the baseline model RT-DETR, while reducing the number of parameters by approximately 37.4%. This fully validates the method’s effectiveness in complex industrial scenarios.

Existing efficient convolutional attention-based super-resolution methods are limited by a constrained receptive field, which results in inadequate detail reconstruction and poor high-frequency feature extraction. To address this issue, we propose a multi-scale asymmetric network, called MSANet, which integrates several strategies to enhance performance and overcome the receptive field limitation. First, we introduce a multi-scale convolutional attention module (MACA), by combining large-kernel attention, multi-scale strategies, and asymmetric structures. Large-kernel attention expands the information aggregation range, thus significantly increasing the receptive field. The multi-scale strategy enhances the effective receptive field by extracting features at multiple levels, thereby improving detail representation. The asymmetric structure allows flexible kernel size configuration, which enables the model boosting the receptive field with minimal computational overhead. Second, to efficiently aggregate features from MACA, we design an enhanced spatial gated block (ESGB), which integrates a gating mechanism and spatial attention. This enables the network to adaptively regulate the information flow, thus improving image quality. Finally, we design a multi-scale asymmetric feature extraction block (MAEB) based on the efficient MetaFormer framework. This ensures both high performance and low resource consumption. Experimental results show that MSANet outperforms most existing efficient super-resolution methods, such as VSCNet, CTE-Net, and demonstrates strong generalization capabilities.

Utilizing Convolutional Neural Networks (CNNs) for image super-resolution reconstruction has emerged as a pivotal technology to enhance image quality. However, existing deep learning models often lack fine details, resulting in unnaturally appearing reconstructed images. Drawing inspiration from the retinal information processing mechanism, we propose a novel super-resolution reconstruction method named RBSRNet. This method employs a lateral inhibition module (LIM) to mimic the regulatory mechanism of horizontal cells, enabling local inhibition and enhancement of shallow extracted features, thereby efficiently extracting high-frequency information. Furthermore, by simulating the information integration mechanism of bipolar cells, we design an enhanced hierarchical feature fusion block (EHFFB) and a hierarchical fusion network (HFN). These components facilitate effective fusion and enhancement of features at different levels, improving the model's representational capacity. Experimental results on four datasets demonstrate that RBSRNet significantly outperforms advanced methods, particularly in texture detail recovery. Specifically, RBSRNet achieves an average PSNR that is 0.26dB higher than IRN on the Urban100 dataset, with reconstructed images exhibiting superior visual effects and richer texture details. The source code is available at https://github.com/Ljjsisr/RBSRNet.