In recent years, computer vision tasks have gradually become a focus and have been widely applied in fields such as intelligent driving, medicine, and remote sensing. However, traditional computer vision tasks lack detailed recognition of category-level content, especially the boundaries of category targets. Semantic segmentation predicts the semantic information of targets by learning annotated category information, thereby separating different categories. However, traditional semantic segmentation requires a large amount of annotation, which leads to high labor costs. Few-shot semantic segmentation aims to segment unannotated categories using only a small number of annotated samples, showing significant advantages in computer vision tasks. This study analyzes papers published in top journals and top conferences over the past five years and points out the existing limitations.

With the increasingly complex electromagnetic environment, wireless communication is facing severe challenges, making modulation recognition of electromagnetic signals, which becomes an important aspect of cognitive radio technology. Deep learning techniques have poor interpretability and little applicability, while traditional identification techniques have limited representation capabilities. In this paper, we propose an intelligent modulation recognition method that combines the advantages of both methods by embedding domain knowledge. In order to enhance classification performance and network interpretability, this technique integrates deep neural networks with high-order information and electromagnetic signal spectrum processes. Based on the RML2018 dataset, our method achieves a 6.31% improvement in modulation recognition accuracy compared to the ResNet method.

Label scarcity poses a critical challenge for data-driven Automatic Modulation Classification (AMC) in realistic electromagnetic environments. To address this limitation, we propose Semi-supervised Multimodal Fusion (SS-MMF), a semi-supervised framework that integrates Mean Teacher consistency learning with multimodal feature fusion for leveraging large-scale unlabeled data. SS-MMF introduces an Attention-Based Multimodal Fusion (ABM-Fusion) module that adaptively integrates time-domain, frequency-domain and time-frequency representations via auxiliary-guided gating, thereby leveraging complementary information while suppressing noise interference. A Modulation-Aware Gated Convolution (MA-GConv) module is introduced to enhance local feature discrimination, while an additive attention mechanism captures long-range dependencies. Extensive experiments demonstrate that SS-MMF consistently outperforms existing methods under limited supervision, achieving improvements of 4.58%, 3.50%, and 2.90% over state-of-the-art semi-supervised approaches on RadioML2016.10a, RadioML2016.10b, and RadioML2018.01a, respectively, with only 10% labeled data, while maintaining stable performance across SNRs and superior accuracy in low-SNR and scarce-label settings.

Self-supervised monocular depth estimation is highly vulnerable to moving objects, which violate the static-scene assumption and cause feature mismatches and occlusions during view synthesis, leading to significant degradation in depth accuracy. To address this, we propose SCFP-Depth, a two-stage framework designed for robust monocular depth estimation in complex scenes. The framework first decouples dynamic and static regions using an adaptive dynamic masking mechanism together with a pretrained optical flow network. A static assumption compensation module (SAC) then generates pseudo-static frame pairs, providing reliable self-supervisory signals for subsequent depth refinement. To further refine depth estimates, we introduce a direction-perception cost volume (DPCV) based on the dual-tree complex wavelet transform (DTCWT), which captures low-frequency semantic structures and high-frequency multi-directional edge details in the frequency domain, thereby strengthening local geometric priors and fine-grained matching. In addition, we design a Fourier-based global information modulation module (GIMM) that integrates multiscale complex filters with a gated fusion mechanism to modulate the fused features and geometric priors, adaptively boosting critical frequency responses while suppressing spatial noise. Experimental results on KITTI, Cityscapes, and Make3D demonstrate that our method consistently outperforms existing representative approaches in both static and dynamic regions, achieving state-of-the-art performance. The code is available at https://github.com/Sixcoins6-maker/SCFP-Depth.

Video surveillance systems are expanding continuously. In this context, the ability to quickly and precisely detect abnormal behaviors from large-scale video data has become essential for intelligent security. Nevertheless, current approaches tend to emphasize video-level discrimination, often neglecting temporal dependency modeling and multi-scale feature fusion. This neglect limits their capacity to detect both short-term abrupt anomalies and events that unfold over long durations. To tackle this limitation, we propose a CLIP-driven Contrastive Adaptive Multi-Scale Temporal learning framework (CLIP-CAMT) for weakly supervised video anomaly detection. Our framework leverages CLIP-generated pseudo-labels as supervision and incorporates an Attention-Enhanced Multi-Scale Temporal Encoder (AEMTE) to capture temporal representations. A crossscale contrastive learning strategy is further employed to improve feature consistency and discriminability across different temporal scales, which in turn enhances anomaly localization accuracy. Additionally, we propose an adaptive multi-scale fusion mechanism that dynamically adjusts fusion weights based on the contribution of each scale, producing a more precise anomaly score curve. Experiments on UCF-Crime and XD-Violence show that CLIP-CAMT outperforms existing approaches on both benchmarks.

Lip-reading recognition is a technology that converts the visual information of a speaker’s lip movements into corresponding textual content. It has broad applications in fields such as national defense, healthcare, and public safety, and holds significant academic value. In recent years, with the rapid advancement of deep learning, lip-reading technology has made notable progress, achieving numerous innovative and breakthrough results. This paper proposes a novel lip-reading recognition architecture that integrates a Residual Network (ResNet) with a Temporal Convolutional Network (TCN), and introduces a simple yet highly effective attention mechanism—Simple Attention Module (SimAM). The key components of the proposed approach are as follows: (1) Feature Extraction: ResNet is employed to extract spatial features from lip images. By introducing residual connections into conventional convolutional neural networks, ResNet effectively alleviates information loss and mitigates the vanishing gradient problem, allowing for more efficient utilization of deep-layer features. (2) SimAM: Traditional attention mechanisms often focus on enhancing features along either the spatial or channel dimension, limiting their ability to learn complex, multi-dimensional attention weights, and typically incurring high computational costs. To address these limitations, SimAM is incorporated. It leverages a spatial suppression mechanism to compute attention weights for each neuron, requiring no additional parameters, while simultaneously attending to both spatial and channel dimensions. (3) Temporal Modeling: TCN is adopted for sequence modeling, applying convolutional operations along the temporal axis. Unlike recurrent networks, TCN enables parallel computation, captures long-range dependencies effectively, and offers a simpler architecture with faster training and greater stability—particularly well-suited for large-scale lip-reading datasets. To validate the effectiveness of the proposed model, experiments were conducted on the largest publicly available lip-reading dataset, LRW, which features diverse pronunciation scenarios and a large number of samples. Comparative experiments with various state-of-the-art architectures demonstrate that the proposed model achieves significant improvements in both recognition accuracy and computational efficiency.

Video, which is distinct from a simple image, encompasses both spatial and temporal dimensions. In the spatial dimension, it contains various visual elements similar to those in static images. However, the addition of the temporal dimension makes it far more complex. It includes static image features such as color, texture, shape, and edge information that are crucial for identifying objects within each frame. Moreover, motion features play a significant role as they describe the movement of objects over time, including velocity, acceleration, and direction of movement. Additionally, external features like lighting conditions, background clutter, and occlusions also affect the overall nature of the video.As an important branch within the broad field of video understanding, human action recognition has attracted widespread attention from the research community and industries alike. The ability to accurately recognize human actions in videos has numerous applications, ranging from surveillance systems to human computer interaction, sports analysis, and entertainment.At present, there are three mainstream methods for processing video data, especially for action recognition: C3D, two stream network, and (2+1) D Net.SlowFast is a typical variant of C3D.The core of SlowFast is to process videos using two channels. These two channels are named Slow pathway and Fast pathway respectively. Compared with Fast pathway, Slow pathway has a relatively lower frame rate but has a greater number of channels. Slow pathway is used to capture semantic information in space, that is, Slow pathway captures the relatively static information in the video.While Fast pathway has a higher frame rate but a smaller number of channels. This greatly reduces the computational complexity of Fast. At the same time, it weakens Fast's ability to model spatial information and makes it pay more attention to information with obvious changes in the temporal dimension.Slow pathway and Fast pathway do not exist independently. The information fusion between the two is unidirectional information fusion. The two achieve information fusion through multiple lateral connections. And the direction of the lateral connections is from Fast to Slow. This means that Fast pathway will not receive any information about Slow pathway. This will undoubtedly lose some semantic information that describes space. We believe that adopting a more effective feature fusion method can further improve the recognition accuracy.Based on the well-known two branch network SlowFast, this paper introduces a significant improvement. Specifically, we propose an enhanced SlowFast network named ESL Net. A key innovation in this network is the addition of an improved 3D feature fusion module. This module is designed to make the most of the temporal information available in the video for effective feature fusion. It employs temporal and spatial attention mechanisms to precisely identify the most significant parts of the features. By analyzing the temporal information, it can also determine the crucial elements between dual - temporal features. Extensive experiments have demonstrated that our proposed method is highly effective when applied to the UCF-101 dataset and the HMDB51 dataset, showing superior performance compared to existing methods especially SlowFast Network in terms of accuracy and robustness in human action recognition tasks.

As a major branch of video understanding, human action recognition has become a popular research topic in the field of computer vision and has a wide range of applications in many areas. To address the problems of high parameter consumption and weak spatiotemporal modeling capabilities in existing action recognition methods, this study proposes a lightweight dual-branch convolutional network called SlowFast-Light Net. Inspired and influenced by the renowned two-branch SlowFast network proposed by the expert Kaiming He, this study adopts a lightweight two-branch network design, which is an improvement based on the SlowFast network. The network significantly reduces parameter consumption by introducing a lightweight feature extraction network and accelerating the model convergence speed. This study conducts experimental verification on the UCF101 and HMDB51 datasets, achieving an action recognition accuracy of 93.80% and 80.00%, respectively, on the two test sets. The experimental results showed that the model proposed in this study achieved a recognition accuracy comparable to that of the original model with a considerably lower number of parameters.

With the rapid advancement of autonomous driving technology, the demand for efficient and accurate 3D object detection systems has significantly increased. To achieve effective cross-modal 3D object detection, existing methods often enhance LiDAR-based detection performance by leveraging the rich semantic information provided by the camera modality, such as improving depth perception and semantic understanding through image branches. However, these approaches typically rely on simplistic feature fusion strategies to align camera and LiDAR features, failing to fully account for the geometric and semantic differences between the two modalities. This limitation adversely impacts detection accuracy and system robustness. To address this issue, this paper proposes an innovative multimodal detection framework, OIGD-MDFusion. The framework incorporates an Object-Intrinsic Geometry Distillation (OIGD) module to improve the alignment accuracy between camera and LiDAR features and employs a Multi-Decoupled Attention Module to enhance the expressive capabilities of both point cloud and image modalities. Experimental results on the nuScenes benchmark demonstrate that OIGD-MDFusion achieves significant improvements in detection accuracy while maintaining outstanding computational efficiency. Notably, compared to existing methods, our model exhibits strong competitiveness on the nuScenes test benchmark, achieving a 72.9% mAP and 75.6% NDS.

Video anomaly detection distinguishes between normal and abnormal events based on differences in feature representations. However, most existing methods are still limited in their video representation capabilities, and there is an imbalance in the data volumes of normal and abnormal events. To address these issues, we propose a multi-instance weakly supervised training framework that efficiently optimizes task-specific discriminative representations using video-level label information. The core components of the framework include: (1) an axial attention-enhanced feature extractor, which integrates global video information and maintains long-range spatial dependencies; (2) a fused BLSTM module, which captures spatiotemporal features before and after abnormal events, automatically focusing on anomalous regions within frames and their contextual information; (3) hard negative feature set, which selects the highest-scoring segments from normal video clips, dynamically updated using a cross-period sampling strategy to ensure the classifier is continuously optimized with the latest challenging samples. Experimental results show that our method outperforms many existing approaches in terms of accuracy and false positive rate, achieving a frame-level accuracy of 94.32

Automatic Modulation Classification (AMC) is vital for adaptive wireless communication, yet it faces challenges in complex environments, including insufficient feature extraction, feature redundancy, and high interclass similarity among modulation schemes. To address these limitations, this paper proposes the Multiscale Complex Convolution Spatiotemporal Attention Network (MCCSAN). In this work, we propose three key innovations tailored for AMC tasks: a multiscale complex convolutional module that directly processes raw I/Q sequences, preserving critical phase and amplitude information while extracting diverse signal features. A spatiotemporal attention mechanism dynamically weights temporal steps and feature channels to suppress redundancy and enhance discriminative feature focus. A combined loss function integrating cross-entropy and center loss improves intraclass compactness and interclass separability. Evaluated on the RML2018.01A dataset and RML2016.10A across SNR levels from −6 dB to 12 dB, MCCSAN achieves a state-of-the-art classification accuracy of 97.03% (peak) and an average accuracy improvement of 3.98% over leading methods. The study confirms that integrating complex-domain processing with spatiotemporal attention significantly enhances AMC performance.

Monocular depth estimation is a critical research area in the field of environmental perception. Self-supervised monocular depth estimation faces challenges such as structural inconsistencies in depth and blurred boundaries due to the lack of precise supervisory signals. To address these issues, we propose a monocular depth estimation architecture based on incremental knowledge distillation and feature decoupling self-distillation, called IKFD-Depth. First, we introduce an incremental knowledge distillation (IKD) method to alleviate the problem of depth information inconsistency. Incremental knowledge is extracted by mining implicit depth information, which decouples the teacher's depth map pseudo-labels into explicit depth values and implicit high-dimensional probability distributions that integrate semantic and spatial information. This implicit information serves as a supplementary signal, providing better supervision for the student network to learn depth consistency. Additionally, we propose a Self-Decoupled Enhancement Module (SDEM), which generates reliable structural features as proxy supervision and employs a Feature Decoupling self-Distillation strategy to extract precise semantic and boundary information from the proxy features. This module effectively assists the depth network in recovering complete scene structures, enabling self-learning-based enhancement and mitigating the performance limitations of conventional teacher networks. Extensive experiments demonstrate that IKFD-Depth achieves the best performance among lightweight models on KITTI and Make3D. On the KITTI dataset, the compact IKFD-Depth (2.8M parameters) achieves average improvements of approximately 4% over Lite-Mono (3.1M parameters) in terms of Abs Rel and Sq Rel.

In autonomous driving environment perception systems, accurate detection of road targets such as vehicles and pedestrians is a key technology for ensuring driving safety. However, existing methods still face challenges such as missed detection of small targets and insufficient stability in detecting occluded targets in complex scenarios. To address these issues, this study proposes an improved YOLOv11-based model. First, an Adaptive Spatial Feature Fusion (ASFF) mechanism is integrated into the detection head, optimizing multi-scale feature fusion via dynamic weight allocation. This enhances the model's performance in detecting objects across different sizes. Next, to boost the detection of small targets, a four-head detection structure is employed, with the addition of high-resolution detection heads improving the retention of small target features. Moreover, a loss function based on the Normalized Gaussian Wasserstein Distance (NWD) is introduced, where bounding boxes are modeled as Gaussian distributions, and their similarity is calculated through distributional similarity, which helps mitigate positional deviations for small targets. Experimental results on both the KITTI and BDD100K datasets show that the improved model not only achieves significant gains in detection accuracy but also maintains its lightweight design. Overall, the proposed approach outperforms existing methods, validating its effectiveness.

Depth estimation plays a vital role in various fields such as autonomous driving, 3D reconstruction, virtual reality, and robotics. Taking self-supervised monocular depth estimation as an example, existing methods commonly enhance multi-scale information transmission through dense cross-layer feature fusion techniques, such as skip connections and channel concatenation. However, these operations often lead to increased feature dimensionality and computational redundancy. Moreover, indiscriminate fusion of low- and high-level features tends to introduce noise, thereby degrading the edge sharpness and geometric consistency of the predicted depth maps. To address these issues, this paper proposes a general-purpose dynamic channel compression framework tailored for feature fusion. A task-oriented channel selection strategy is employed to reduce redundancy while preserving critical information. Specifically, the proposed framework includes a dynamic threshold generation network that determines sample-adaptive retention thresholds based on the distribution of channel-wise standard deviations. Furthermore, a probabilistic L0 gating mechanism is introduced, which relaxes the discrete channel selection problem into a continuous optimization via a hard shrinkage function. The design is attention-free and easily integrable into existing architectures. Experimental results on the KITTI dataset demonstrate that our method achieves a 10%–30% reduction in parameter count across multiple depth estimation models, with accuracy degradation of less than 1%.

Weakly Supervised Video Anomaly Detection (WSVAD) is a critical task in computer vision. It aims to localize and recognize abnormal behaviors using only video-level labels. Without frame-level annotations, it becomes significantly challenging to model temporal dependencies. Given the diversity of abnormal events, it is also difficult to model semantic representations. Recently, the cross-modal pre-trained model Contrastive Language-Image Pretraining (CLIP) has shown a strong ability to align visual and textual information. This provides new opportunities for video anomaly detection. Inspired by CLIP, WSVAD-CLIP is proposed as a framework that uses its cross-modal knowledge to bridge the semantic gap between text and vision. First, the Axial-Graph (AG) Module is introduced. It combines an Axial Transformer and Lite Graph Attention Networks (LiteGAT) to capture global temporal structures and local abnormal correlations. Second, a Text Prompt mechanism is designed. It fuses a learnable prompt with a knowledge-enhanced prompt to improve the semantic expressiveness of category embeddings. Third, the Abnormal Visual-Guided Text Prompt (AVGTP) mechanism is proposed to aggregate anomalous visual context for adaptively refining textual representations. Extensive experiments on UCF-Crime and XD-Violence datasets show that WSVAD-CLIP notably outperforms existing methods in coarse-grained anomaly detection. It also achieves superior performance in fine-grained anomaly recognition tasks, validating its effectiveness and generalizability.

Although self-supervised depth estimation models based on transformers have achieved success, lightweight depth prediction networks exhibit a particularly pronounced issue with depth prediction blurriness at object boundaries compared to standard depth prediction networks. We found that this problem arises from the token dimension constraints, which limit the precise representation of semantic and spatial information. To address this challenge, we introduce a lightweight monocular self-supervised depth estimation network, RENA-Depth, which leverages convolutional neural network neighborhood attention-guided recursive Transformers to enhance depth estimation precision. Specifically, the design begins with the introduction of neighborhood adaptive attention (NA), which focuses on local and regional scales. This component adaptively mines latent semantic and spatial information from the neighborhoods of the input features of self-attention. Subsequently, a global feature recursive interaction module was developed to recursively refine the interaction between local and global information, enhancing the representation of semantic and spatial information without a significant increase in parameters. Finally, an attention equilibrium loss is proposed, which motivates richer semantic information representation and clarifies boundary depth by penalizing the orthogonality similarity of attention mechanisms. Extensive evaluations on the Karlsruhe Institute of Technology and Toyota Technological Institute and Make3D datasets have demonstrated that the proposed lightweight self-supervised depth estimation model, RENA-Depth, outperforms the most advanced lightweight depth detection algorithms, confirming its efficacy and innovation in improving depth prediction accuracy.

In the realm of image deraining, traditional CNN-based deep learning deraining systems exhibit efficient expression of local features and strong generalization capabilities. However, their limited local receptive fields and independence from input content hinder their ability to model global features, rendering them less effective in mitigating complex and dynamic long rain streak scenarios. On the other hand, deraining systems based on the Transformer architecture possess robust global feature aggregation capabilities. Yet, their computational complexity increases quadratically with the expansion of the image spatial scale, making them less suitable for high-quality image deraining tasks. However, for ill-posed problems like image deraining, the precise representation of both local and global features has become increasingly pivotal to addressing the multifaceted challenges of rain streak removal. Consequently, we introduce an innovative solution, the CNN and Gated Multi Axial-Sparse Transformer Feature Fusion Network, referred to as CGMAformer. This approach optimizes both architectural paradigms jointly, effectively harnessing their respective strengths for image deraining. Specifically, in the local feature extraction phase based on CNN, we employ the Degradation-aware Mixture of Experts Feature Compensator (DMEFC) for adaptive representation of local spatial rain streak features. In the global feature extraction phase based on the Transformer, we introduce a dual-branch adaptive Gated Multi-Axis Sparse Transformer (GAST) attention mechanism to complement global background spatial features in rainy images. This approach ensures the preservation of global feature integrity while effectively reducing model complexity. Ultimately, through a feature fusion network, we fully exploit the local characteristics of CNN and the self-attention-based global aggregation capabilities of the Transformer for efficient image deraining.

Remote sensing scene classification is a hot research topic in computer vision and it is of great significance to the semantic understanding of remote sensing images. At present, remote sensing scene classification methods based on deep learning occupy a dominant position in this field. However, it suffers from the lack of samples and poor model generalization ability in actual application scenarios. Therefore, this paper proposes a few-shot remote scene classification method based on attention mechanism, and designs a structure of dual-branches similarity measurement. This method is based on the meta-learning training strategy to divide the dataset into tasks. At the meantime, the input images are divided into blocks to preserve the feature distribution in the remote sensing image. Then the lightweight attention module is introduced into the feature extraction network to reduce the risk of overfitting and ensure the acquisition of discriminative features. Finally we add a dual-branches similarity measurement module on the basis of Earth Mover’s Distance to improve the discriminative ability of the classifier. The results show that compared with the classic small-sample learning method, the few-shot remote scene classification method proposed in this paper can significantly improve the classification performance.

Currently, lip recognition technology is a significant research direction in the field of video understanding in computer vision. It aims to recognize the content expressed by the main characters through the dynamic changes of the lips' visual image. With the development of deep learning and enhanced computer performance, lip recognition techniques have evolved to effectively separate the background and target foreground across different scenes, improving uniformity in models and technical routes. However, despite the success of lip recognition with English datasets, the semantic specificity of Chinese words and the scarcity of open-source Chinese lip recognition datasets have generally led to subpar recognition standards. To address these challenges, this paper introduces a novel model that synergizes two-dimensional and three-dimensional networks. Our approach leverages an improved Convolutional 3D (C3D) network to extract spatiotemporal features effectively. Unlike traditional 2D networks, which lack temporal dynamics, and conventional 3D networks that are prone to overfitting due to deep layers, our enhanced C3D network provides a robust foundation for feature extraction. To capture temporal features more efficiently, we integrate the outputs into a Bidirectional Gated Recurrent Unit (Bi-GRU), combined with a Multi-Head Self-Attention mechanism. This fusion allows for better extraction of semantic and syntactic features, overcoming the limitations of normal RNNs in handling long sequences and the non-parallelizable nature of GRU due to sequence dependence. The effectiveness of the proposed model was validated through experiments on our self-compiled Chinese dataset. We compared it with mainstream networks such as ResNet-18, ResNet-34, and the original C3D model lacking the multi-attention mechanism. The analysis of loss function curves and accuracy curves demonstrated that our model achieves significant improvements, effectively addressing the challenges in lip recognition for the Chinese language and setting a new benchmark for performance in this field.