In recent years, self-supervised monocular depth estimation methods have shown significant potential in fields such as autonomous driving, virtual reality, and robotics. However, the existing lightweight approaches often struggle with feature representation because of the limitations imposed by smaller model sizes, leading to issues such as missing scene objects and blurred edge details. To address these challenges, this study proposes a novel lightweight self-supervised monocular depth estimation method. Specifically, this study introduces a Joint Attention-guided Multiscale spatial Feature Fusion (JAMFF) skip connection strategy that effectively bridges the gap in detail and semantics between the encoder and decoder. Additionally, this study employs a Lightweight Residual Cross-dimensional information Interaction module (LRCI) at all encoder—decoder stages. On the one hand, LRCI ensures that the relationships between spatial and channel are fully considered to protect and enhance depth features for small and weak objects or parts of them; on the other hand, the design of cross-dimensional information interaction is a crucial aspect in ensuring the lightweight nature of the model. The experimental results on the KITTI dataset show that the proposed method achieves an accuracy of 0.889, with Absolute Relative Error (ARE), Square Relative Error (SRE), Root Mean Square Error (RMSE), and Root Mean Squared Log Error (RMSLE) of 0.106, 0.789, 4.581, and 0.182, respectively. Additionally, the model parameter size is only 3.3×106, which generates more accurate depth maps while maintaining model lightweighting. The test results on other datasets with complex scenarios and harsh conditions indicate that the algorithm has strong generalization ability.
Medical report generation from images is challenging owing to the low image contrast and small size of abnormal regions, making it difficult to accurately capture abnormal features using visual information alone. Therefore, the introduction of external knowledge to enhance visual representations is a key issue. Additionally, the co-occurrence patterns of abnormal features are complex and cannot be effectively captured from a single instance, making it crucial to leverage similar cases to model such patterns. To address these challenges, a Similar Instance Guidance (SIG) approach that fuses heterogeneous graphs for medical image report generation is presented. This approach consists of two main components: an image feature memory module incorporating heterogeneous graphs and a Similar Instance Feature Fusion (SIFF) module. The image feature memory module incorporating heterogeneous graphs extracts entity relationships from reports and constructs a corresponding heterogeneous graph as a bridge, guiding the model's attention to the abnormal regions of the image, thus enhancing abnormal visual features. SIFF retrieves similar instances and integrates their abnormal visual features, thereby augmenting the representation of abnormal regions, while acquiring a more comprehensive understanding of the abnormal information. To verify the effectiveness of the proposed method, experiments are conducted on both IU X-Ray and MIMIC-CXR datasets. The results indicate that the proposed method outperforms baseline methods on BLEU-based metrics, yielding BLEU-1-BLEU-4 values of 0.539, 0.353, 0.265, and 0.193, respectively, on the IU X-Ray dataset. Additionally, it excels in the METEOR and ROUGE-L metrics, indicating that the proposed method outperforms existing methods in terms of natural language generation metrics as well as the accuracy and completeness of the generated reports.
Micro-gestures are unconscious, subtle movements driven by intrinsic emotions. As they are capable of reflecting individuals' hidden emotional states, they hold significant value in affective computing. Micro-gestures exhibit transient characteristics in the temporal dimension and spatially ambiguous boundaries. Hence, these typical fine-grained behaviors pose substantial challenges for feature extraction by traditional methods. To address these issues, this paper proposes a micro-gesture recognition method based on multi-modal collaborative enhancement that constructs video, skeleton, and text modalities as a complementary representational triplet. This framework addresses the limitations of conventional vision-language models by introducing skeleton modality as a kinematic prior. By integrating visual context and semantic guidance, it establishes a multi-source complementary feature representation system. A dual-level collaborative module, called the Video—Pose Collaborative Module (VPCM), fuses fine-grained video details with global motion information from skeletons, and a cross-temporal attention mechanism is expands feature representations and enhances temporal modeling capability. The Text-Pose Collaborative Module (TPCM) introduces semantic priors from the text modality and adopts a Top-k-based fusion strategy to reinforce the semantic relevance of skeleton features, thereby improving the capture of fine-grained characteristics. To further optimize the multi-modal fusion performance, a two-stage training strategy is designed, in which unimodal encoders are pre-trained first, followed by collaborative learning through lightweight adapters and collaborative modules, thus effectively boosting model accuracy. Results of experiments on mainstream micro-gesture datasets demonstrate that the proposed model achieves a micro-gesture recognition accuracy of 70.40%, outperforming the current best-performing methods.
The surfaces of Printed Circuit Boards (PCB) feature complex textures with defects that are small as well as diverse in type and shape. Typically, accurate detection of small targets necessities the addition of detection heads at smaller scales, which significantly increases the computational cost and reduces the detection speed. To address this issue, this study proposes PCB-Det, a multi-scale feature fusion learning model for small target defect detection on a PCB. Building on the YOLOv8 architecture, PCB-Det replaces the original backbone network with a lightweight backbone, PP-HGNet, and incorporates a GSPPFCSPC module for multi-level feature extraction, expanding the receptive field to enrich feature information. Meanwhile, a Progressive Bidirectional Feature Pyramid Network (Pro-BiFPN) multi-scale feature fusion network is designed to enhance the interactions between adjacent feature layers, optimizing the fusion of shallow detail information and deep semantic information. In addition, the model employs a shared feature branch to reduce the weight of the original detection head and introduces the Wise-IoU loss function, which dynamically adjusts loss weights to accelerate model convergence. The experimental results show that PCB-Det achieves an Average Precision (AP) of 98.1% on PCB_DATASET, representing a 3.5 percentage point improvement over the baseline model, effectively reducing missed detections and false alarms and enhancing the detection capability for small target PCB defects.
In the field of cybersecurity, phishing attacks are becoming increasingly complex and frequent. Traditional phishing detection schemes based on predefined reference templates rely on brand-domain mapping lists, using visual feature matching to identify brand intent and verify domain consistency for explainable detection. While these methods can counter zero-day phishing attacks, they face scalability challenges due to the need for continuous updates to reference lists to cover emerging brands, leading to high maintenance costs. To address these, the paper proposes Phish-RAGLLM, a novel reference-based phishing detection scheme leveraging Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG). By reframing traditional visual problems into language tasks, Phish-RAGLLM eliminates reliance on predefined reference lists, utilizing LLMs' extensive brand knowledge while enhancing generation capabilities through RAG integration with external brand knowledge bases. This approach effectively mitigates LLM hallucination issues and improves detection precision and robustness. Experimental results demonstrate that, compared to the current state-of-the-art model PhishLLM, Phish-RAGLLM—using GPT-3.5-turbo-instruct as the main LLM—balances model performance, inference cost, and knowledge base completeness, achieving 5.88% increase in F1 score and a 12.5% improvement in operational efficiency. Moreover, it shows strong robustness against dataset variations and prompt injection attacks. Based on the characteristics of LLM, Phish-RAGLLM exhibits good adaptability to multilingual phishing websites, effectively detecting phishing webpages in different linguistic contexts. Furthermore, real-world evaluations reveal that Phish-RAGLLM has broader detection capabilities than VirusTotal (a threat intelligence source), further validating its feasibility and effectiveness.
