Discriminative Representation Research Articles

Node classification in complex networks based on multi-view debiased contrastive learning

AbstractIn complex networks, contrastive learning has emerged as a crucial technique for acquiring discriminative representations from graph data. Maximizing the similarity among relevant sample pairs while minimizing that among irrelevant pairs is pivotal in contrastive learning. Therefore, careful consideration must be given to the design of sample pairs in contrastive learning. However, existing node-level self-supervised contrastive learning often treats the enhanced representation of a central node as a positive sample, while considering representations of all other nodes as negative samples. This approach can lead to conflicts in downstream tasks on some graph data, as nodes of the same class are treated as negatives during learning. Precision in sample pair design is essential for enhancing the performance of contrastive learning. To address this issue, this paper introduces a negative sample debiased sampling contrastive learning (NDSCL), specifically tailored for node classification tasks. In particular, this method integrates contrastive learning with semi-supervised learning. A trained classifier assigns pseudo-labels to unlabeled data, and debiased sampling is applied to negative samples. Unlike other methods that focus on negative sample selection, NDSCL also addresses the imbalance in pseudo-label distribution by employing debiasing techniques. Finally, in conjunction with diffusion augmentation, the model is provided with diverse views as inputs to maximize the retention of underlying semantic information. Experimental results demonstrate that the proposed model significantly outperforms baseline models in node-level classification tasks across multiple network datasets. Moreover, the model not only enhances accuracy but also improves computational speed and memory requirements for handling large-scale graph data structures.

Multimodal MRI segmentation of key structures for microvascular decompression via knowledge-driven mutual distillation and topological constraints.

Microvascular decompression (MVD) is a widely used neurosurgical intervention for the treatment of cranial nerves compression. Segmentation of MVD-related structures, including the brainstem, nerves, arteries, and veins, is critical for preoperative planning and intraoperative decision-making. Automatically segmenting structures related to MVD is still challenging for current methods due to the limited information from a single modality and the complex topology of vessels and nerves. Considering that it is hard to distinguish MVD-related structures, especially for nerve and vessels with similar topology, we design a multimodal segmentation network with a shared encoder-dual decoder structure and propose a clinical knowledge-driven distillation scheme, allowing reliable knowledge transferred from each decoder to the other. Besides, we introduce a class-wise contrastive module to learn the discriminative representations by maximizing the distance among classes across modalities. Then, a projected topological loss based on persistent homology is proposed to constrain topological continuity. We evaluate the performance of our method on in-house dataset consisting of 100 paired HR-T2WI and 3D TOF-MRA volumes. Experiments indicate that our model outperforms the SOTA in DSC by 1.9% for artery, 3.3% for vein and 0.5% for nerve. Visualization results show our method attains improved continuity and less breakage, which is also consistent with intraoperative images. Our method can comprehensively extract the distinct features from multimodal data to segment the MVD-related key structures and preserve the topological continuity, allowing surgeons precisely perceiving the patient-specific target anatomy and substantially reducing the workload of surgeons in the preoperative planning stage. Our resources will be publicly available at https://github.com/JaronTu/Multimodal_MVD_Seg .

Deep Spectral Clustering With Constrained Laplacian Rank.

Spectral clustering (SC) is a well-performed and prevalent technique for data processing and analysis, which has attracted significant attention in the field of clustering. While the scalability and generalization ability of this method make it prohibitive for the large-scale dataset and the out-of-sample-extension problem. In this work, we propose a new efficient deep clustering architecture based on SC, named deep SC (DSC) with constrained Laplacian rank (DSCCLR). DSCCLR develops a self-adaptive affinity matrix with a clustering-friendly structure by constraining the Laplacian rank, which greatly mines the intrinsic relationships. Meanwhile, by introducing a simple fully connected network with an orthogonality constraint on the last layer, DSCCLR learns discriminative representations in a short training time. The proposed method has the following salient properties: 1) it overcomes limited generalization ability and scalability of the existing DSC methods; 2) it explores the intrinsic relationship between samples in the affinity matrix, which maintains the latent manifold of data as much as possible; and 3) it alleviates the complexity of eigendecomposition via a simple but effective fully connected network. The extensive empirical results demonstrate the superiorities of DSCCLR over other 17 clustering methods.

TransEBUS: The interpretation of endobronchial ultrasound image using hybrid transformer for differentiating malignant and benign mediastinal lesions

The purpose of this study is to establish a deep learning automatic assistance diagnosis system for benign and malignant classification of mediastinal lesions in endobronchial ultrasound (EBUS) images. EBUS images are in the form of video and contain multiple imaging modes. Different imaging modes and different frames can reflect the different characteristics of lesions. Compared with previous studies, the proposed model can efficiently extract and integrate the spatiotemporal relationships between different modes and does not require manual selection of representative frames. In recent years, Vision Transformer has received much attention in the field of computer vision. Combined with convolutional neural networks, hybrid transformers can also perform well on small datasets. This study designed a novel deep learning architecture based on hybrid transformer called TransEBUS. By adding learnable parameters in the temporal dimension, TransEBUS was able to extract spatiotemporal features from insufficient data. In addition, we designed a two-stream module to integrate information from three different imaging modes of EBUS. Furthermore, we applied contrastive learning when training TransEBUS, enabling it to learn discriminative representation of benign and malignant mediastinal lesions. The results show that TransEBUS achieved a diagnostic accuracy of 82% and an area under the curve of 0.8812 in the test dataset, outperforming other methods. It also shows that several models can improve performance by incorporating two-stream module. Our proposed system has shown its potential to help physicians distinguishing benign and malignant mediastinal lesions, thereby ensuring the accuracy of EBUS examination.

Uncertainty-guided label correction with wavelet-transformed discriminative representation enhancement

Label noises, categorized into closed-set noise and open-set noise, are prevalent in real-world scenarios and can seriously hinder the generalization ability of models. Identifying noise is challenging because noisy samples closely resemble true positives. Existing approaches often assume a single noise source, oversimplify closed-set noise, or treat open-set noise as toxic and eliminate it, resulting in limited practical effects. To address these issues, we present a novel approach named uncertainty-guided label correction with wavelet-transformed discriminative representation enhancement (Ultra), designed to mitigate the effects of mixed noise. Specifically, our approach considers a more practical noise setting. To achieve robust mixed-noise identification, we initially look into a learnable wavelet filter for obtaining discriminative features and filtering spurious cues automatically at the representation level. Subsequently, we introduce a two-fold uncertainty estimation to stably locate noise within the corrupted supervised signal level. These insights pave the way for a simple yet potent label correction technique, enabling comprehensive utilization of open-set noise, which can be rendered non-toxic in a specific manner, in contrast to harmful closed-set noise. Experimental validation on datasets with synthetic mixed noise, web noise corruption, and a real-world dataset confirms the effectiveness and generality of Ultra. Furthermore, our approach enhances the application of efficient techniques (e.g., supervised contrastive learning) within label noise scenarios.

Clustering-Guided Twin Contrastive Learning for Endomicroscopy Image Classification.

Learning better representations is essential in medical image analysis for computer-aided diagnosis. However, learning discriminative semantic features is a major challenge due to the lack of large-scale well-annotated datasets. Thus, how can we learn a well-structured categorizable embedding space in limited-scale and unlabeled datasets? In this paper, we proposed a novel clustering-guided twin-contrastive learning framework (CTCL) that learns the discriminative representations of probe-based confocal laser endomicroscopy (pCLE) images for gastrointestinal (GI) tumor classification. Compared with traditional contrastive learning, in which only two randomly augmented views of the same instance are considered, the proposed CTCL aligns more semantically related and class-consistent samples by clustering, which improved intra-class tightness and inter-class variability to produce more informative representations. Furthermore, based on the inherent properties of CLE (geometric invariance and intrinsic noise), we proposed to regard CLE images with any angle rotation and CLE images with different noises as the same instance, respectively, for increased variability and diversity of samples. By optimizing CTCL in an end-to-end expectation-maximization framework, comprehensive experimental results demonstrated that CTCL-based visual representations achieved competitive performance on each downstream task as well as more robustness and transferability compared with existing state-of-the-art SSL and supervised methods. Notably, CTCL achieved 75.60%/78.45% and 64.12%/77.37% top-1 accuracy on the linear evaluation protocol and few-shot classification downstream tasks, respectively, which outperformed the previous best results by 1.27%/1.63% and 0.5%/3%, respectively. The proposed method holds great potential to assist pathologists in achieving an automated, fast, and high-precision diagnosis of GI tumors and accurately determining different stages of tumor development based on CLE images.

A contrastive learning approach for enhanced robustness for strengthening federated intelligence in internet of visual things

As the Internet of Things (IoT) continues to grow, edge intelligence has developed as a favorable paradigm to enable effective and instantaneous processing of data at the network's edge. The deployment of an edge intelligence approach in untrusted environments exposes them to adversarial attacks, posing a substantial challenge to the reliability and confidentiality of the learned models especially in federated settings. Among these attacks, the white-box attacks are the most dangerous and challenging to defend in IoT systems. This study proposes a novel approach to robustify the federated edge intelligence technique by integrating the power of contrastive learning to enhance the model's ability to distinguish between clean and adversarial examples, thereby improving its robustness. Our robustification approach is based on two lines of defense. First, an elegant federated contrastive pretraining is introduced to robustify the model by adversarially training it using a label-free representation attack. Second, we introduce a novel Consistency Regularized Triplet (CRT) loss function that encourages the regularization of the federated edge intelligence model to learn discriminative representations while minimizing the influence of adversarial perturbations. The empirical evaluation of the Food-101 and CIFAR-100 datasets revealed that the proposed method outperformed the state-of-the-art robust methods on both standard accuracy and robustness accuracy.

Mitigating biases in long-tailed recognition via semantic-guided feature transfer

Dealing with significant class imbalance poses a significant challenge in various real-world applications, particularly when the accurate classification and generalization of minority classes are of crucial. One key factor that results in model biases is the inadequate representation of intra-class diversity in samples from tail classes. To tackle this challenge, transfer learning strategies have emerged as a viable approach to promote the data distribution of tail classes by transferring feature learned from head classes. However, existing methods often assume that transferability exists between any head class and tail class, and randomly transfer information during training, which can be problematic when the head and tail classes have uncorrelated semantics. To tackle this problem, we propose a novel feature transfer strategy to promote the learning of discriminative representations for tail classes. Firstly, we compute the semantic similarity in a hybrid manner by combining local visual features and global semantic features. Secondly, we utilize the Gaussian mixture model to represent each class with multiple and anisotropic prototypes, which helps us fit the actual distribution accurately. We then carry out distribution-aware head-to-tail feature transfer, leveraging an optimization objective that can derive a computationally efficient closed-form upper bound. This procedure mitigates representation biases in the feature space and achieves state-of-the-art results on CIFAR100-LT, ImageNet-LT, and iNaturalist long-tailed benchmarks.

Enhancing UAV tracking: a focus on discriminative representations using contrastive instances

Enhancing UAV tracking: a focus on discriminative representations using contrastive instances

SaliencyCut: Augmenting plausible anomalies for anomaly detection

Anomaly detection under the open-set scenario is a challenging task that requires learning discriminative features to detect anomalies that were even unseen during training. As a cheap yet effective approach, data augmentation has been widely used to create pseudo anomalies for better training of such models. Recent wisdom of augmentation methods focuses on generating random pseudo instances that may lead to a mixture of augmented instances with seen anomalies, or out of the typical range of anomalies. To address this issue, we propose a novel saliency-guided data augmentation method, SaliencyCut, to produce pseudo but more common anomalies that tend to stay in the plausible range of anomalies. Furthermore, we deploy a two-head learning strategy consisting of normal and anomaly learning heads to learn the anomaly score of each sample. Theoretical analyses show that this mechanism offers a more tractable and tighter lower bound of the data log-likelihood. We then design a novel patch-wise residual module in the anomaly learning head to extract and assess anomaly features from each sample, facilitating the learning of discriminative representations of anomaly instances. Extensive experiments conducted on six real-world anomaly detection datasets demonstrate the superiority of our method to competing methods under various settings. Codes are available at: https://github.com/yjnanan/SaliencyCut.

Adjustable patch and feature prior token-based transformer for weakly supervised semantic segmentation

Weakly supervised semantic segmentation is a challenging task, utilizing only low-cost weak supervision to produce pixel-level predictions. Existing transformer-based methods for weakly supervised semantic segmentation have some limitations: (1) Fixing patch size might destroy the structured semantics, which is unfriendly to objects of different scales, and (2) Ignoring the prior features when using multi-head attention mechanisms might lead to inaccurate segmentation localization. To tackle these issues, we proposed an effective transformer framework coupled with the adjustable patch and prior feature tokens, termed as APFPformer, in which an adjustable patch module is developed to split the image according to the area of the salient object for preserving the structured semantics in patches. A prior feature token module is devised to exploit the edge and texture information as prior tokens, ensuring the gain of discriminative representation. Additionally, a single-stage scheme is applied to reduce the computation and rapid segmentation process. Our experiments demonstrate the superiority of our approach over early methods, gaining competitive mean Intersection-over-Union scores of 67.9% on the PASCAL VOC2012 dataset and 40.2% on the MS COCO dataset.

RamanCluster: A deep clustering-based framework for unsupervised Raman spectral identification of pathogenic bacteria

Raman spectroscopy serves as a powerful and reliable tool for the characterization of pathogenic bacteria. The integration of Raman spectroscopy with artificial intelligence techniques to rapidly identify pathogenic bacteria has become paramount for expediting disease diagnosis. However, the development of prevailing supervised artificial intelligence algorithms is still constrained by costly and limited well-annotated Raman spectroscopy datasets. Furthermore, tackling various high-dimensional and intricate Raman spectra of pathogenic bacteria in the absence of annotations remains a formidable challenge. In this paper, we propose a concise and efficient deep clustering-based framework (RamanCluster) to achieve accurate and robust unsupervised Raman spectral identification of pathogenic bacteria without the need for any annotated data. RamanCluster is composed of a novel representation learning module and a machine learning-based clustering module, systematically enabling the extraction of robust discriminative representations and unsupervised Raman spectral identification of pathogenic bacteria. The extensive experimental results show that RamanCluster has achieved high accuracy on both Bacteria-4 and Bacteria-6, with ACC values of 77 % and 74.1 %, NMI values of 75 % and 73 %, as well as AMI values of 74.6 % and 72.6 %, respectively. Furthermore, compared with other state-of-the-art methods, RamanCluster exhibits the superior accuracy on handling various complicated pathogenic bacterial Raman spectroscopy datasets, including situations with strong noise and a wide variety of pathogenic bacterial species. Additionally, RamanCluster also demonstrates commendable robustness in these challenging scenarios. In short, RamanCluster has a promising prospect in accelerating the development of low-cost and widely applicable disease diagnosis in clinical medicine.

ProtoMGAE: Prototype-Aware Masked Graph Auto-Encoder for Graph Representation Learning

Graph self-supervised representation learning has gained considerable attention and demonstrated remarkable efficacy in extracting meaningful representations from graphs, particularly in the absence of labeled data. Two representative methods in this domain are graph auto-encoding and graph contrastive learning. However, the former methods primarily focus on global structures, potentially overlooking some fine-grained information during reconstruction. The latter methods emphasize node similarity across correlated views in the embedding space, potentially neglecting the inherent global graph information in the original input space. Moreover, handling incomplete graphs in real-world scenarios, where original features are unavailable for certain nodes, poses challenges for both types of methods. To alleviate these limitations, we integrate masked graph auto-encoding and prototype-aware graph contrastive learning into a unified model to learn node representations in graphs. In our method, we begin by masking a portion of node features and utilize a specific decoding strategy to reconstruct the masked information. This process facilitates the recovery of graphs from a global or macro level and enables handling incomplete graphs easily. Moreover, we treat the masked graph and the original one as a pair of contrasting views, enforcing the alignment and uniformity between their corresponding node representations at a local or micro level. Last, to capture cluster structures from a meso level and learn more discriminative representations, we introduce a prototype-aware clustering consistency loss that is jointly optimized with the preceding two complementary objectives. Extensive experiments conducted on several datasets demonstrate that the proposed method achieves significantly better or competitive performance on downstream tasks, especially for graph clustering, compared with the state-of-the-art methods, showcasing its superiority in enhancing graph representation learning.

Cluster structure augmented deep nonnegative matrix factorization with low-rank tensor learning

Clustering approaches based on deep nonnegative matrix factorization have received significant attention because it can learn hierarchical semantics from data in a layer-wise manner. However, latent representation learning and clustering are two separate processes in these models, leading to information loss and sub-optimal clusterings. Furthermore, if data contain complex structure or noise, the pre-defined graph in data space will not be precise enough for graph-regularized deep nonnegative matrix factorization models. Moreover, most of them maintain the similarity graph at the top layer, neglecting the information covered by other layers. In this paper, Cluster Structure Augmented Deep Nonnegative Matrix Factorization with Low-rank Tensor Learning is proposed. First, as representations at different layers cover different data abstractions, a learning mechanism is leveraged to generate multiple local similarity graphs based on representations from different layers, thus maintaining diversity across layers. Then, a low-rank third-order tensor is constructed by stacking these graphs to capture the high-order consistency across layers, thus capturing the intrinsic property of data. Third, clustering is integrated into the framework, allowing the cluster structure to facilitate multi-layer matrix factorization and graph structure learning, thereby generating discriminative representations. Experimental results on several datasets demonstrate that the proposed model outperforms several state-of-the-art methods.

Edge‐guided representation learning for underwater object detection

AbstractUnderwater object detection (UOD) is crucial for marine economic development, environmental protection, and the planet's sustainable development. The main challenges of this task arise from low‐contrast, small objects, and mimicry of aquatic organisms. The key to addressing these challenges is to focus the model on obtaining more discriminative information. The authors observe that the edges of underwater objects are highly unique and can be distinguished from low‐contrast or mimicry environments based on their edges. Motivated by this observation, an Edge‐guided Representation Learning Network, termed ERL‐Net is proposed, that aims to achieve discriminative representation learning and aggregation under the guidance of edge cues. Firstly, an edge‐guided attention module is introduced to model the explicit boundary information, which generates more discriminative features. Secondly, a hierarchical feature aggregation module is proposed to aggregate the multi‐scale discriminative features by regrouping them into three levels, effectively aggregating global and local information for locating and recognising underwater objects. Finally, a wide and asymmetric receptive field block is proposed to enable features to have a wider receptive field, allowing the model to focus on smaller object information. Comprehensive experiments on three challenging underwater datasets show that our method achieves superior performance on the UOD task.

Mining negative samples on contrastive learning via curricular weighting strategy

Contrastive learning, which pulls positive pairs closer and pushes away negative pairs, has remarkably propelled the development of self-supervised representation learning. Previous studies either neglected negative sample selection, resulting in suboptimal performance, or emphasized hard negative samples from the beginning of training, potentially leading to convergence issues. Drawing inspiration from curriculum learning, we find that learning with negative samples ranging from easy to hard improves both model performance and convergence rate. Therefore, we propose a dynamic negative sample weighting strategy for contrastive learning. Specifically, we design a loss function that adaptively adjusts the weights assigned to negative samples based on the model's performance. Initially, the loss prioritizes easy samples, but as training advances, it shifts focus to hard samples, enabling the model to learn more discriminative representations. Furthermore, to prevent an undue emphasis on false negative samples during later stages, which probably results in trivial solutions, we apply L2 regularization on the weights of hard negative samples. Extensive qualitative and quantitative experiments demonstrate the effectiveness of the proposed weighting strategy. The ablation study confirms both the reasonableness of the curriculum and the effectiveness of the regularization.

DualGAD: Dual-bootstrapped self-supervised learning for graph anomaly detection

Graph anomaly detection (GAD) is an emerging and essential research field for discovering anomalous individuals (e.g., nodes or edges) that deviate significantly from the normal majority in an attributed graph. Unlike other anomaly types (e.g., image, text) with independence, it is not trivial to capture inherent and distinctive anomaly patterns as graph anomalies usually exist more complex relational interaction with a overwhelming amount of normal majority, which induces the inconsistency in feature and structure space. Recently, some studies have applied graph self-supervised learning on the GAD task and surge a new climax. Despite their success, the sub-optimal performance is only achieved because the imbalance nature of anomaly problem significantly dilutes anomalous information and causes the shortage of effective supervision signals. To address these challenges, we propose a dual-bootstrapped self-supervised approach, namely DualGAD, that consists of one generative module and one contrastive module. Specifically, we first sample a local subgraph for each target node and employ two mask strategies on subgraphs to break the short-range connection and reduce redundancy spread from majority neighbors. The generative module is equipped with the reconstruction objective to model the surrounding context of masked subgraphs, which learn discriminative representations by capturing the inconsistency patterns at both feature and structure space. Considering the local information is easily been over-emphasized, we elaborately tailor a novel cluster-guided contrastive learning module without relying on positive/negative pairs to learn the intrinsic anomaly property by aligning the perturbation distributions across views. Finally, the two self-supervised modules are seamlessly integrated and bootstrap each other for learning the discriminative representations to distinguish anomalies and normal majority. Extensive experiments demonstrate the superiority of DualGAD on the node-level anomaly detection task.

MGCNSS: miRNA-disease association prediction with multi-layer graph convolution and distance-based negative sample selection strategy.

Identifying disease-associated microRNAs (miRNAs) could help understand the deep mechanism of diseases, which promotes the development of new medicine. Recently, network-based approaches have been widely proposed for inferring the potential associations between miRNAs and diseases. However, these approaches ignore the importance of different relations in meta-paths when learning the embeddings of miRNAs and diseases. Besides, they pay little attention to screening out reliable negative samples which is crucial for improving the prediction accuracy. In this study, we propose a novel approach named MGCNSS with the multi-layer graph convolution and high-quality negative sample selection strategy. Specifically, MGCNSS first constructs a comprehensive heterogeneous network by integrating miRNA and disease similarity networks coupled with their known association relationships. Then, we employ the multi-layer graph convolution to automatically capture the meta-path relations with different lengths in the heterogeneous network and learn the discriminative representations of miRNAs and diseases. After that, MGCNSS establishes a highly reliable negative sample set from the unlabeled sample set with the negative distance-based sample selection strategy. Finally, we train MGCNSS under an unsupervised learning manner and predict the potential associations between miRNAs and diseases. The experimental results fully demonstrate that MGCNSS outperforms all baseline methods on both balanced and imbalanced datasets. More importantly, we conduct case studies on colon neoplasms and esophageal neoplasms, further confirming the ability of MGCNSS to detect potential candidate miRNAs. The source code is publicly available on GitHub https://github.com/15136943622/MGCNSS/tree/master.

Federated Momentum Contrastive Clustering

Self-supervised representation learning and deep clustering are mutually beneficial to learn high-quality representations and cluster data simultaneously in centralized settings. However, it is not always feasible to gather large amounts of data at a central entity, considering data privacy requirements and computational resources. Federated Learning (FL) has been developed successfully to aggregate a global model while training on distributed local data, respecting the data privacy of edge devices. However, most FL research effort focuses on supervised learning algorithms. A fully unsupervised federated clustering scheme has not been considered in the existing literature. We present federated momentum contrastive clustering (FedMCC), a generic federated clustering framework that can not only cluster data automatically but also extract discriminative representations training from distributed local data over multiple users. In FedMCC, we demonstrate a two-stage federated learning paradigm where the first stage aims to learn differentiable instance embeddings and the second stage accounts for clustering data automatically. The experimental results show that FedMCC not only achieves superior clustering performance but also outperforms several existing federated self-supervised methods for linear evaluation and semi-supervised learning tasks. Additionally, FedMCC can easily be adapted to ordinary centralized clustering through what we call momentum contrastive clustering (MCC). We show that MCC achieves state-of-the-art clustering accuracy results in certain datasets such as STL-10 and ImageNet-10. We also present a method to reduce the memory footprint of our clustering schemes.

Learning Multi-Scale Video-Text Correspondence for Weakly Supervised Temporal Article Gronding

Weakly Supervised temporal Article Grounding (WSAG) is a challenging and practical task in video understanding. Specifically, given a video and a relevant article, whose sentences are at different semantic scales, WSAG aims to localize corresponding video segments for all “groundable” sentences. Compared to other grounding tasks, e.g., localizing one target segment with respect to a given sentence query, WSAG confronts an essential obstacle rooted in the intricate multi-scale information inherent within both textual and visual modalities. Existing methods overlook the modeling and alignment of such structured information present in multi-scale video segments and hierarchical textual content. To this end, we propose a Multi-Scale Video-Text Correspondence Learning (MVTCL) framework, which enhances the grounding performance in complex scenes by modeling multi-scale semantic correspondence both within and between modalities. Specifically, MVTCL initially aggregates video content spanning distinct temporal scales and leverages hierarchical textual relationships in both temporal and semantic dimensions via a semantic calibration module. Then multi-scale contrastive learning module is introduced to generate more discriminative representations by selecting typical contexts and performing inter-video contrastive learning. Through the multi-scale semantic calibration architecture and supervision design, our method achieves new state-of-the-art performance on existing WSAG benchmarks.