Zheng Guan

Effectively capturing and utilizing the features and complementary potential between different feature layers of an image is of great significance for accurately locating salient objects and preserving their detailed contours. Aiming at the problem of generalization ability caused by the fact that most existing RGB-T saliency detection models input the extracted image features directly into the aggregation modules, and use a simple recursive structure to locate the saliency object, a new method saliency enhancement and global awareness for RGB-T salient object detection is proposed. Firstly, the pre-trained model is used to extract the original features of the image. Secondly, a noise reduction module (NRM) is proposed, which acts as a bridge between encoding blocks and cross-modal decoding blocks to purify the features of salient objects. The high-level semantic guidance block (HSGB) is proposed to combine the high-level semantic information of different modes to preserve the location information of salient objects. Finally, a multi-modal interaction block (MMIB) is designed to retain the location information of salient objects and guide the aggregation of multi-level features to obtain the final salient map. Experimental results on VT821, VT1000 and VT5000 datasets show that the MAE index of the proposed method reaches 3.0%, 1.8% and 3.0%, respectively, which is superior to most existing methods.

With its well-designed network architecture, the deep learning-based infrared and visible image fusion (IVIF) method shows its efficiency and effectiveness by realizing a fine feature extraction and fusion mechanism. However, disparities in cross-modal features often result in an imbalance between texture details and contextual information, causing detailed features to be overshadowed by prevailing contextual information. To tackle this issue, this study introduces PIDFusion, a fusion model driven by a PID controller, designed to dynamically optimize cross-modal feature fusion deviations. The core of PIDFusion is the dynamic adaptation capability of the PID controller, which facilitates real-time corrections for deviations encountered during the fusion process, thereby maintaining a harmonious balance between texture details and contextual information. Additionally, we introduced the Cyclic Self-Supervised Feature Refinement (CSSFR), which under the constraint of self-supervised loss, minimizes redundant information within the feature flow and ensures the preservation of salient feature through the cyclic input of decoupled features. Concurrently, we developed the Iterative Attention Module (IAM), utilizing the unique gating mechanism of LSTM to capture feature changes across successive iterations, thereby driving the model to cultivate more discriminative feature representations. Extensive experiments revealed that PIDFusion outperforms SOTA methods in terms of both efficiency and cost-effectiveness, through static statistics and high-level vision tasks. Our code is available at https://github.com/wang-x-1997/PIDFusion.

E-learning systems, with their flexibility across time and location, further highlight the importance of consumer electronics in education. Knowledge tracing (KT) is a key part, which depicts the state of students’ knowledge state and predicts students’ future performance. However, due to the sparsity of data in real scenarios, the adequacy of state modeling and the accuracy of prediction are affected. To address the above problem, we propose a method (FTCKT) that fuses interaction transition features and cross-attention relations to monitor the state of students’ knowledge by synergistically modeling global semantics and local semantics. Specifically, we employ a directed graph-based global information modeling module that aims to thoroughly extract exercise transition features, further refining the completeness of knowledge state modeling. Furthermore, we also designed a cross-domain attention fusion module to facilitate the integration of complementary information. The fusion module captures the long-term intent of the target exercise by modeling the dependencies of sequence domain, while synergistically modeling with the relational unit of graph domain to fully mine and integrate the long-term dependencies and interaction information among the interaction sequences. The experimental results show that dual-domain collaborative modeling can effectively extract interaction features, enhancing the model’s state modeling and prediction performance.

Knowledge tracing (KT) plays a crucial role in personalized e-learning systems, aiming to model the current knowledge state by tracing the learner’s historical learning records to predict the learner’s future performance. However, most existing deep KT models do not take into account the correlation of questions and knowledge concepts, leading to inaccurate representations of students’ knowledge states. Additionally, obtaining relationships between concepts through manual tagging is not only prohibitively expensive, but also create subjective bias in the knowledge states representation. To address these challenges and explore implicit connections between questions and concepts, we propose a memory-driven attentive knowledge tracing (MAKT) model. The memory module in MAKT mitigates the subjective impact of manually labeling concepts by capturing information from learning interactions, fuzzily modeling the relationships between questions and concepts, and dynamically updating them. To further account for the complexity of learning behavior, we effectively decouple the impact of the question relevance and the learner’s forgetting effects using a linear bias attention mechanism, which enhances the interpretability of the model. We conducted extensive experiments on five publicly available datasets and showed that MAKT significantly outperforms previous knowledge tracing methods in predicting student performance.

Infrared and visible image fusion (IVIF) aims to integrate complementary information between sensors and generate information-rich high-quality images. However, current methods mainly concentrate on the fusion of the source features from the sensors, ignoring the feature information mismatch caused by the property of the sensors, which results in redundant or even invalid information. To tackle the above challenges, this paper developed an end-to-end model based on the Retinex Decomposition Model (RDM), called RDMFuse, which utilizes a hierarchical feature process to alleviate the fusion performance degradation caused by the feature-level mismatch. Specifically, as infrared images only provide an overview of the intrinsic properties of the scene, we first use RDM to decouple visible images into a reflectance component containing intrinsic properties and an illumination component containing illumination information. Then, the contrast texture module (CTM) and the intrinsic fusion function are designed for the property of the intrinsic feature, which complements each other to aggregate the intrinsic information of the source images at a smaller cost and brings the fused image more comprehensive scene information. Besides, the illumination-adaptive module implements illumination component optimization in a self-supervised way to make the fused image with an appropriate intensity distribution. It is worth noting that this mechanism implicitly improves the entropy quality of the image to improve the image degradation problem caused by environmental factors, especially in the case of a dark environment. Numerous experiments have demonstrated the effectiveness and robustness of the RDMFuse and the superiority of generalization in high-level vision tasks due to the improved discriminability of the fused image to the captured scene. The code is available at https://github.com/wang-x-1997/RDMFuse.