To address the problems of complex backgrounds， significant variations in lesion scale， and subtle differences between adjacent categories in multi-level recognition of wheat stripe rust， a Multi-Scale feature fusion and Attention-enhanced Residual Network （MSA-ResNet） was proposed. First， residual connections were employed to alleviate the gradient vanishing during deep network training and improve the feature representation capability of the model. Second， a top-down multi-scale feature fusion strategy was adopted to enhance the consistency of features at different scales through semantic alignment， thereby improving the model’s adaptability to lesion size variations and morphological complexity. Finally， a Convolutional Block Attention Module （CBAM）， integrating channel and spatial attention mechanisms， was introduced to enhance feature responses in key lesion regions and suppress interference from background elements such as weeds and soil. Experimental results on the Yellow-Rust-19 public dataset and a self-built dataset show that MSA-ResNet achieves the accuracies of 95.67% and 97.17%， respectively， outperforming EfficientNet-CBAM by 1.67 and 0.84 percentage points， respectively. These results indicate that the proposed MSA-ResNet better meets practical application requirements and provides a superior solution for multi-level recognition of wheat stripe rust under field conditions.
To address the semantic ambiguity and difficulty in candidate selection caused by the diverse results generated by Large Language Models （LLMs） in intent detection and slot filling tasks， a joint intent-slot recognition method based on prompt-guided generation and tri-mutual attention mechanism was proposed， named PGT-Joint （Prompt-guided Generation and Tri-mutual attention for Joint intent-slot recognition）. First， the prompts were constructed based on domain knowledge to guide the LLM in generating multiple candidate intent-slot combinations. Second， a tri-mutual attention mechanism was designed to model the high-order semantic relationships among user questions， candidate intents， and slots explicitly. Finally， a contrastive learning ranking model was used to filter and optimize the candidate results. Experimental results on a Mini/Micro LED domain question-answering dataset show that compared with current mainstream models except Global-Locally Graph Interaction Network （GL-GIN）， such as BERT-Intent and BERT-Slot， PGT-Joint improves the intent detection accuracy by 0.4 to 3.8 percentage points and the slot filling F1 score by 4.1 to 6.6 percentage points. Notably， the most significant gain is in overall accuracy， with an improvement of 5.2 to 10.2 percentage points. In addition， ablation experimental results verify the key roles of prompt generation and tri-mutual attention mechanism in performance improvement， indicating that PGT-Joint can enhance semantic understanding capabilities in domain-specific dialogue systems effectively.
Deep learning-based image classification algorithms usually rely on huge amounts of training data. However， it is often difficult to obtain sufficient large-scale high-quality labeled samples in real scenarios. Aiming at the problem of insufficient generalization ability of classification models in few-shot scenarios， a few-shot image classification method based on contrast learning was proposed. Firstly， global contrast learning was added as an auxiliary target in training to enable the feature extraction network to obtain richer information from instances. Then， the query samples were split into patches and used to calculate the local contrast loss， thereby promoting the model to gain the ability to infer the global thing the local things. Finally， saliency detection was used to mix the important regions of the query samples， and complex samples were constructed， so as to improve the model generalization ability. Experimental results of 5-way 1-shot and 5-way 5-shot image classification tasks on two public datasets， miniImageNet and tieredImageNet， show that compared to the few-shot learning baseline model， Meta-Baseline， the proposed method improves the classification accuracy by 5.97 and 4.25 percentage points respectively on miniImageNet， and by 3.86 and 2.84 percentage points respectively on tieredImageNet. Besides， the classification accuracy of the proposed method on miniImageNet is improved by 1.02 and 0.72 percentage points respectively compared to that of DFR （Disentangled Feature Representation） model. It can be seen that the proposed method improves the accuracy of few-shot image classification effectively with good generalization ability.
Aiming at the problems that the existing self-supervised contrastive tasks fail to make full use of the rich semantic information in the original data and lack universality， a Session-based Recommendation with Graph Auxiliary Learning （SR-GAL） model was proposed. Firstly， an encoding channel with Representation Consistency （RC） was introduced on the basis of Graph Neural Network （GNN） to mine more valuable self-supervised signals from the original data. Secondly， in order to make full use of these self-supervised signals， two auxiliary tasks， predictive one and constraint one， that were closely related to the target task were designed. Finally， a simple and GNN model-unrelated auxiliary learning framework was developed to unify the two auxiliary tasks with the recommendation task in order to improve the recommendation performance of the GNN model. Compared with the suboptimal comparison model CGSNet （Contrastive Graph Self-attention Network）， on Diginetica dataset， the proposed model has the Precision P@20 and Mean Reciprocal Rank MRR@20 increased by 0.58% and 1.61%； on Tmall dataset， the proposed model has the P@20 and MRR@20 increased by 12.65% and 8.41% respectively， verifying the effectiveness of the model. Experimental results on multiple real datasets show that SR-GAL model outperforms advanced models and has good extensibility as well as universality.
Aiming at the problem that the Multi-Layer Perceptron （MLP） architecture can not capture the co-occurrence relationship in the context of session sequence， a session-based recommendation model based on Graph Co-occurrence Enhanced MLP （GCE-MLP） was proposed. Firstly， the sequential dependency of the session sequence was captured by the MLP architecture， and at the same time， the co-occurrence relationship in the sequence context was obtained through the co-occurrence relationship learning layer， and the session representation was obtained through the information fusion module. Secondly， a specific feature selection layer was designed to amplify the diversity of input features of different relation learning layers. Finally， the representation learning of sessional interest was further enhanced by maximizing the mutual information between two relational representations via a noise contrastive task. Experimental results on multiple real datasets show that the recommendation performance of the GCE-MLP is better than those of the current mainstream models， which verifies the effectiveness of GCE-MLP. Compared with the optimal MLP architecture model FMLP-Rec（Filter-enhanced MLP for Recommendation）， GCE-MLP achieves the P@20 of 54.08% and the MRR@20 of 18.87% for Diginetica dataset， which are respectively increased by about 2.14 and 1.43 percentage points； GCE-MLP achieves the P@20 of 71.77% and the MRR@20 of 31.78% for Yoochoose dataset， which are respectively increased by about 0.48 and 1.77 percentage points.
The existing monocular depth estimation methods often use image semantic information to obtain depth， and ignore another important cue — defocus blur. At the same time， the defocus blur based depth estimation methods usually take the focal stack or gradient information as input， and do not consider the characteristics of the small variation of blur between image layers of the focal stack and the blur ambiguity on both sides of the focal plane. Aiming at the deficiencies of the existing focal stack depth estimation methods， a lightweight network based on three-dimensional convolution was proposed. Firstly， a Three-Dimensional perception module was designed to roughly extract the blur information of the focal stack. Secondly， the extracted information was concatenated with the difference features of the focal stack RGB channels output by a channel difference module to construct a focus volume that was able to identify the blur ambiguity patterns. Finally， a multi-scale three-dimensional convolution was used to predict the depth. Experimental results show that compared with methods such as All in Focus Depth Network （AiFDepthNet）， the proposed method achieves the best on seven indicators such as Mean Absolute Error （MAE） on DefocusNet dataset， and the best on four indicators as well as the suboptimal on three indicators on NYU Depth V2 dataset； at the same time， the lightweight design reduces the inference time of the proposed method by 43.92% to 70.20% and 47.91% to 77.01% on two datasets respectively. The above verifies that the proposed method can effectively improve the accuracy and inference speed of focal stack depth estimation.
The variable-length address is one of the important research content in the field of future network. Aiming at the low efficiency of traditional routing lookup algorithms for variable-length address， an efficient routing lookup algorithm suitable for variable-length addresses based on balanced binary tree — AVL （Adelson-Velskii and Landis） tree and Bloom filter， namely AVL-Bloom algorithm， was proposed. Firstly， multiple off-chip hash tables were used to separately store route entries with the same number of prefix bits and their next-hop information in view of the flexible and unbounded characteristics of the variable-length address. Meanwhile， the on-chip Bloom filter was utilized for speeding up the search for route prefixes that were likely to match. Secondly， in order to solve the problem that the routing lookup algorithms based on hash technology need multiple hash comparisons when searching for the route with the longest prefix， the AVL tree technology was introduced， that was， the Bloom filter and hash table of each group of route prefix set were organized through AVL tree， so as to optimize the query order of route prefix length and reduce the number of hash calculations and then decrease the search time. Finally， comparative experiments of the proposed algorithm with the traditional routing lookup algorithms such as METrie （Multi-Entrance-Trie） and COBF （Controlled prefix and One-hashing Bloom Filter） were conducted on three different variable-length address datasets. Experimental results show that the search speed of AVL-Bloom algorithm is significantly faster than those of METrie and COBF algorithms， and the query time is reduced by nearly 83% and 64% respectively. At the same time， AVL-Bloom algorithm can maintain stable search performance under the condition of large change in routing table entries， and is suitable for routing lookup and forwarding with variable-length addresses.
KGE(Knowledge Graph Embedding） maps entities and relationships into a low-dimensional continuous vector space， uses machine learning methods to implement relational data applications， such as knowledge analysis， reasoning， and completion. Taking ConvE （Convolution Embedding） as a representative， CNN （Convolutional Neural Network） is applied to knowledge graph embedding to capture the interactive information of entities and relationships， but the ability of the standard convolutional to capture feature interaction information is insufficient， and its feature expression ability is low. Aiming at the problem of insufficient feature interaction ability， an improved Inception structure was proposed， based on which a knowledge graph embedding model named InceE was constructed. Firstly， hybrid dilated convolution replaced standard convolution to improve the ability to capture feature interaction information. Secondly， the residual network structure was used to reduce the loss of feature information. The experiments were carried out on the datasets Kinship， FB15k， WN18 to verify the effectiveness of link prediction by InceE. Compared with ArcE and QuatRE models on the Kinship and FB15k datasets， the Hit@1 of InceE increased by 1.6 and 1.5 percentage points； compared with ConvE on the three datasets， the Hit@1 of InceE increased by 6.3， 20.8， and 1.0 percentage points. The experimental results show that InceE has a stronger ability to capture feature interactive information.
The modeling of user-item interaction patterns is an important task for personalized recommendation. Many recommendation systems are based on the assumption that there is a linear relationship between users and items， and ignore the complexity and non-linearity of interaction between real and historical items， as a result， these systems cannot capture the complex decision-making process of users. Therefore， a more expressive top-N recommendation system’s item similarity factor model solution was combined with the multi-layer perceptron approach， to effectively model the higher-order relationships between items and capture more complex user decisions. The combination effect was verified on the three datasets of MovieLens， Foursquare and ratings_Digital_Music； and compared with the benchmark methods such as MLP （Multi-Layer Perception）， Factored Item Similarity Model （FISM）， DeepICF （Deep Item-based Collaborative Filtering） and ItemKNN （Item-based K-Nearest Neighbors）， the results demonstrate that the proposed method has significant improvement in recommendation performance.
