Wanzhi Wen

Transformer architectures have demonstrated remarkable efficacy in image restoration; however, their quadratic self‐attention complexity imposes severe computational constraints for deployment on resource‐limited platforms. We investigate an alternative pathway within the transformer family itself: replacing quadratic attention with linear attention mechanisms while preserving the architectural strengths of spatial‐wise and channel‐wise feature modelling. We introduce Mixed Linear Transformer (MLT) blocks, each integrating multi‐head spatial‐wise linear self‐attention and multi‐head channel‐wise linear self‐attention, to construct a lightweight student network. To compensate for the representational capacity reduction inherent to linearization, we propose a Dual Feature‐based Knowledge Distillation (DFKD) framework that operates in an offline manner: a high‐capacity transformer‐based teacher is first pre‐trained via supervised learning and subsequently frozen, while the student learns to replicate both spatial‐wise and channel‐wise attention patterns through dedicated feature alignment losses. Extensive evaluation across multiple image restoration benchmarks demonstrates that the distilled MLT‐based model, operating with merely 1.45 M parameters and 12.8 GFLOPs, achieves competitive or superior performance relative to state‐of‐the‐art approaches.

Benefiting from the powerful nonlinear fitting ability of neural networks, deep learning‐based methods have gradually emerged as the dominate solutions for single image dehazing. However, supervised learning–based methods require paired image samples for training. To address this, an unsupervised Vision Mamba with Contrastive Regularization network (VMCR) is proposed. The network is designed based on the DisentGAN framework, and its main module is the Vision Mamba. This module performs very competitively compared to transformers, while maintaining linear time complexity and constant memory complexity with respect to the input size. Furthermore, a contrastive regularization method based on contrastive learning is proposed to enhance the reconstruction capabilities of the network and achieve superior dehazing results. Our VMCR‐Net outperforms state‐of‐the‐art unsupervised image dehazing methods, as evidenced by experimental results on several benchmarks. This research successfully proposes an enhanced unsupervised image dehazing approach, overcoming the limitations of existing methods and achieving superior dehazing performance.

Speech produced in helium–oxygen (heliox) environments in deep saturation diving exhibits pronounced spectral shifts and temporal distortions, which severely degrade automatic speech recognition (ASR) systems trained on normal-air corpora. Existing studies often adopt a restoration-then-recognition paradigm by training waveform mapping networks on paired heliox/air recordings. However, in realistic low-resource data collection, paired recordings are typically obtained by independent re-reading and are therefore not strictly time-aligned, which makes regression-style restoration more sensitive to pairing errors and increases the risk of front-end distortions. This paper proposes a robust recognition framework for heliox speech, termed PRL-DAS (Physics-informed Resampling and LoRA with Duration-Adaptive Speed). The framework consists of a physics-inspired linear resampling warm start (PhysSpeed), parameter-efficient Low-Rank Adaptation (LoRA), and duration-adaptive speed (DAS) inference enhancement. Specifically, we first apply physics-motivated linear resampling as a coarse warm start, and then perform mixed-domain LoRA fine-tuning of a Whisper foundation model to absorb residual non-linear differences. On a corpus of 1048 paired Chinese heliox utterances under leave-one-speaker-out (LOSO) evaluation, using Whisper-Medium as the base model, PhysSpeed followed by mixed-domain LoRA reduces the overall character error rate (CER) from 49.33% with PhysSpeed preprocessing only to 25.79%, while also improving performance on the normal domain. Furthermore, the full PRL-DAS framework applies Soft-DAS, a lightweight smooth schedule motivated by duration-dependent variation in the optimal resampling factor, and further reduces the overall CER to 24.37% without additional training cost.

This study introduces a rapid and non-destructive method for detecting multiple chemical residues in pure cotton fabrics, utilizing hyperspectral imaging, discrete wavelet transform (DWT), and a one-dimensional convolutional neural network (1D-CNN) to address limitations of traditional techniques, which have low efficiency and destructive nature. Initially, spectral data were collected from pure cotton fabric samples using a hyperspectral imaging system operating within the 400–1000 nm wavelength range , with each sample yielding one averaged spectrum from 100 pixels. The dataset was split 75:25 for training (9,000 spectra) and testing (3,000 spectra), using stratified sample-wise partitioning to prevent spectral leakage. Subsequently, DWT was applied for data dimensionality reduction, eliminating redundant information while retaining critical spectral characteristics. An optimized 1D-CNN model was developed to automatically extract spectral features and classify multiple residues, including formaldehyde, 4-aminobiphenyl, nonylphenol, melamine, and ethanol. The experimental findings demonstrated that the DWT preserved high information entropy (13.02) even after dimensionality reduction to 75 features. Furthermore, its computational efficiency significantly surpassed that of competitive adaptive reweighted sampling and kernel principal component analysis. In conjunction with DWT for dimensionality reduction, the 1D-CNN model achieved an overall classification accuracy of 99.76% and an F1 score of 0.99, exceeding the performance of the support vector machine and random forest models by 2.03% and 38.53%, respectively. This methodology facilitates non-destructive and rapid detection, achieving over 50% increase in detection speed while maintaining high accuracy in identifying formaldehyde residue. It represents an effective technological solution for ensuring textile quality and safety. Future work will focus on designing lightweight network architectures, implementing data augmentation strategies, and expanding the model’s applicability to detect multiple pollutants, aiming to enhance its practical deployment and generalization capability.

Cross-project defect prediction technique is a hot topic in the field of software defect research because of the huge difference in data distribution between source project and target project. Software defect prediction technique usually first extracts software project features and then trains prediction models based on various classifiers. However, traditional features lack sufficient semantic information of source code resulting in poor performance of the prediction models. To construct more accurate prediction models based on the semantic information, we propose a cross-project defect prediction framework named BSLDP, which extracts semantic information of source code files through a bidirectional long and short-term memory network with self-attention mechanism. In particular, we provide semantic extractor named ALC to extract source code semantics based on source code files, and propose classification algorithm based on the semantic information of source project and target project, namely BSL, to build a prediction model. Furthermore, we propose an equal meshing mechanism that ALC generates semantic information on small fragments by dividing the numerical token vector to further improve the performance of the proposed model. We evaluated the performance of the proposed model on a publicly available PROMISE dataset. Compared with the four state-of-the-art methods, the experimental results indicate that on average BSLDP improves the performance of cross-project defect prediction in terms of F1 by 14.2%, 34.6%, 32.2% and 23.6%, respectively.

Software defect prediction techniques can help software developers find software defects as soon as possible. It can also reduce the cost of software development. This technique usually predicts the target project through the entire source project. However, the data distribution difference between the entire source project and the target project is generally large, so the software defect prediction accuracy is not high. we propose a cross-project software defect prediction technique based on class code similarity CCS-CPDP. Firstly, this technique converts the code set extracted by AST(Abstract Syntax Tree) into a vector set through the DTI (Doc2Bow and TF-IDF) strategy; Secondly, the similarity will be calculated between the vector set of target projects and training projects; Finally, according to the principle of the majority decision subordinate category in KNN, the number of most similar class instances of the training project is determined, the source project is refined by selecting the class instance, and then software defects are predicted and evaluated. We compared CCS-CPDP with softawre defect prediction methods based on four traditional classification models (KNN, Random Forest, Naive Bayes, and Logistic Regression). Experimental results show that CCS-CPDP can improve the effectiveness of CPDP in terms of recall and f1-score.

Source code comments can improve the efficiency of software development and maintenance. However, due to the heterogeneity of natural language and program language, the quality of code comments is not so high. So, this paper proposes a novel method Code2tree, which is based on the encoder-decoder model to automatically generate Java code comments. Code2tree firstly converts Java source code into abstract syntax tree (AST) sequences, and then the AST sequences are encoded by GRU encoder to solve the long sequence learning dependency problem. Finally, the attention mechanism is introduced in the decoding stage, and the quality of the code comment is improved by increasing the weight of the key information. We use the open dataset java-small to train the model and verify the effectiveness of Code2tree based on common-used indicators BLEU and F1-Score.