Enhancing Lightweight IRSR Models via Knowledge Distillation with Structural and Spectral Losses

Due to the limitation of hardware, Infrared (IR) image has low-resolution (LR) and poor visual quality. Infrared image super-resolution (SR) is a good solution for this problem. However, the conventional SR methods have some drawbacks. Firstly, the trained dictionary is an unstructured dictionary, which may lead to worse results. Secondly, the representation of the image is too simple to effectively represent image. To resolve these problems, in this paper, firstly, the sparse dictionary is introduced into the IR image SR to get better results. Secondly, nonsubsampled contour let transform (NSCT) is employed in the proposed method to obtain a better representation of IR image. The experiment results indicate that the subjective visual effect and objective evaluation are acquired excellent performance in the proposed method. Besides, this method is superior to other methods in the paper.

Recent studies have shown that Deep Learning (DL) algorithms can significantly improve Super Resolution (SR) performance. Single image SR is useful in producing High Resolution (HR) images from their Low Resolution (LR) counterparts. The motivation for SR is the potential to assist algorithms such as object detection, localization, and classification. Insufficient work has been conducted using Generative Adversarial Networks (GANs) for SR on infrared (IR) images despite its promising ability to increase object detection accuracy by extracting more precise features from a given image. This work adopts the idea of a relativistic GAN that utilizes Residual in Residual Dense blocks (RRDBs) for feature ex- traction, a novel residual image addition, and a Pixel Transposed Convolutional Layer (PixelTCL) for up-sampling. Recent work has validated the use of GANs for Visible Light (VL) images, making them a strong candidate. The inclusion of these components produce more realistic and natural features while also receiving superior metric values.

Infrared imaging has the advantage of all-weather working ability. Due to the limitation of the hardware and the high cost, the resolution of infrared image (IR) is very low. To improve the resolution of IR images, this paper exploits super-resolution (SR) method for IR images. A new SR framework by using random forests is proposed in this paper. Existing methods adopts single regression model for SR. However, which single regression model tends to overfit training data, and would lead to a poor performance. Furthermore, the existing methods are not suitable for real-time system due to the heavy time consuming. To resolve this problem, an ensemble regression model, i.e. random forests rather than single regression model is adopted in this paper. In addition, to achieve better results multi-regression models rather than a single regression model are trained on the clustered training data. Moreover, the features used in many SR methods cannot extract features on diagonal orientation. To resolve this problem, we adopt a second order derivative filter, which can extract features on diagonal orientation. The experimental results demonstrate the availability of the proposed method.

Infrared images hold significant application value in fields such as military reconnaissance, security surveillance, and medical diagnosis. However, issues like low resolution, high noise, and complex degradation characteristics severely hinder their practical application effectiveness. This paper introduces an infrared super-resolution reconstruction algorithm based on a random degradation model and Generative Adversarial Networks (GANs), addressing the diversity of infrared image degradation processes. The primary contribution lies in explicitly modeling key degradation parameters of infrared images (such as blur kernel and noise distribution) using a random degradation model, generating diverse low-resolution to high-resolution image pairs, and significantly enhancing the model’s generalization ability for complex degradations. Experiments on the Airo infrared dataset and a self-built infrared dataset demonstrate that when the resolution is increased by a factor of 2, 4, and 8, the 4× resolution reconstructed images exhibit notable advantages in terms of texture detail and noise suppression. Especially in 4× super-resolution reconstruction, compared to three typical deep learning algorithms, our algorithm improves the Peak Signal-to-Noise Ratio (PSNR) by 3.046 dB, 1.8489 dB, and 0.2108 dB, respectively, and the Structural Similarity Index (SSIM) by 0.0387 (4.76%), 0.0287 (3.48%), and 0.0131 (1.56%), respectively, with perceptual similarity decreasing by 0.2465, 0.13344, and 0.0514 (lower values indicate better perceptual quality). Subjective visual assessments further validate the algorithm’s significant advantages in noise reduction and weak texture restoration. This study proposes an infrared image super-resolution reconstruction method based on random degradation modeling, which holds significant theoretical and practical value in complex infrared degradation scenarios.

Infrared image super-resolution via transformed self-similarity

Accurate monitoring of wheat phenology is critical for ensuring wheat production. Recent advances in deep learning have enabled the automated detection of wheat phenology in the field. In particular, deep learning models using multi-temporal image series have addressed the challenge of low accuracy in models that only use spatial features by incorporating dynamic aspects of the wheat growth process. However, utilizing multi-temporal image series introduces challenges such as model parameter redundancy, complex inference processes, and difficulties in real-time deployment. To address these issues, this study presents an optimization method for deriving wheat phenology from single-temporal images (WPDSI) that combines knowledge distillation and multi-layer attention transfer. The proposed approach employs knowledge distillation. In this framework, a teacher model extracts spatiotemporal features from multi-temporal image-series and generates soft labels to guide a student model trained on single-temporal images. This reduces model complexity and input data requirements. Multi-layer attention transfer allows the student model to inherit feature representations from multiple layers of the teacher model. This enhances its ability to capture key phenological characteristics and supports interpretability through attention mechanisms. The proposed method achieves an overall accuracy (OA) of 0.927, comparable to models trained on multi-temporal image series. Furthermore, the model demonstrates strong generalization on unseen datasets, enhancing real-time performance and computational efficiency while maintaining high accuracy, providing a practical solution for deriving wheat phenology in the field. The dataset is available at https://github.com/phenology-detection/WPDSI.

Security is one of the most important things in our daily lives and the security business is a big industry. Security cameras are set in many areas in order to keep us safe. When a crime occurs, we often hear that police analyze the security camera footage. It takes long time and it is often the case that the footage is not helpful in the investigation at all. The reason is very simple. Many of the security camera images do not have sufficient resolution. Currently there are many super resolution (SR) technologies. In detective dramas technicians solve crimes using SR technologies turning blurry images taken with a security camera into sharp high resolution ones with a click of a button. However, in real life it is not so easy. In this paper we discuss the current issues and limitations of the SR technologies.

Surface defect detection is a core component for ensuring the quality of multi-wedge belt products, yet traditional detection methods based on Convolutional Neural Networks (CNNs) suffer from weak global feature extraction and limited multi-scale feature fusion, severely impacting detection efficiency. Meanwhile, the constrained computational resources in industrial scenarios pose critical challenges to balancing model accuracy and speed. To address these issues, we propose a lightweight defect detection model, LCT-YOLO (Lightweight CNN-Transformer YOLO), which integrates the technical advantages of CNN and Transformer while meeting the lightweight deployment requirements of industrial applications. Specifically, a dynamic channel selection feature extraction network is constructed by introducing the partial channel selection module CSP_PTB to enhance the model’s capability of capturing long-distance semantic information. A lightweight multi-scale feature fusion network is developed by designing a multi-scale path aggregation network and integrating the dynamic upsampling module Dysample in the model’s neck, effectively addressing challenges from feature scale variations. A lightweight shared convolution detection head is introduced to tackle the high computational cost of existing model detection heads, optimizing performance and reducing model complexity through parameter sharing strategies. Finally, model pruning is employed to remove redundant parameters for compression, and knowledge distillation is utilized to compensate for accuracy loss. Experimental results show that the algorithm achieves an mAP50 of 94.3% on the dataset, with 0.8 M parameters and 2.0 GFLOPs computation, outperforming the baseline model by 0.9% in mAP50, reducing parameters by 74.2%, computation by 75.3%, and featuring a model size of only 1.9 M, making it highly suitable for edge device deployment. The algorithm surpasses other advanced methods in both accuracy and speed, not only meeting the dual requirements of detection precision and real-time performance in industrial scenarios but also being successfully applied to practical production, providing robust technical support for defect detection and product quality control.

Forest fires pose a significant threat to ecosystems, property, and human life, making their early and accurate detection crucial for effective intervention. This study presents a novel, lightweight approach to real-time forest fire detection that is optimized for resource-constrained devices like drones. The method integrates multi-task knowledge distillation, transferring knowledge from a high-performance DenseNet201 teacher model that was trained on a hierarchically structured wildfire dataset. The dataset comprised primary classes (fire vs. non-fire) and detailed subclasses that account for confounding elements such as smoke, fog, and reflections. The novelty of this approach lies in leveraging knowledge distillation to transfer the deeper insights learned by the DenseNet201 teacher model—specifically, the auxiliary task of recognizing the confounding elements responsible for false positives—into a lightweight student model, enabling it to achieve a similar robustness without the need for complex architectures. Using this distilled knowledge, we trained a MobileNetV3-based student model, which was designed to operate efficiently in real-time while maintaining a low computational overhead. To address the challenge of false positives caused by visually similar non-fire elements, we introduced the Confounding Element Specificity (CES) metric. This novel metric, made possible by the hierarchical structure of the wildfire dataset, is unique in its focus on evaluating how well the model distinguishes actual fires from the confounding elements that typically result in false positives within the negative class. The proposed approach outperformed the baseline methods—including single-task learning and direct multi-task learning—achieving a primary accuracy of 93.36%, an F1-score of 91.57%, and a higher MES score, demonstrating its enhanced robustness and reliability in diverse environmental conditions. This work bridges the gap between advanced deep learning techniques and practical, scalable solutions for environmental monitoring. Future research will focus on integrating multi-modal data and developing adaptive distillation techniques to further enhance the model’s performance in real-time applications.

This study explored the integration of super-resolution (SR) image reconstruction technology into environmental art design, addressing key challenges and highlighting its potential to enhance visual quality, design efficiency, ecological awareness, user satisfaction, and cultural heritage preservation. By examining the principles and models of SR technology, the study demonstrated its ability to significantly improve the aesthetic and functional standards of environmental art, contributing to more sustainable and culturally rich designs. The analysis and practical applications presented underscored the significance of SR technology in advancing both technological and artistic progress. Future developments in SR technology promise to further transform environmental art design, creating more dynamic, sustainable, and visually compelling environments.

Spectroscopic analysis and modeling face challenges due to the high-dimensional and heterogeneous nature of spectral data, compounded by complex noise interference. Traditional spectral processing methods, including feature engineering, spectral preprocessing, and data augmentation, are limited by subjective bias, spurious feature artifacts, and physical inconsistencies. Knowledge distillation (KD) facilitates the optimization of raw spectral information by transferring dark knowledge to a student. To achieve further model compression, (self-) KD (self-KD) involves discarding the large teacher model and its redundant branches. However, this may hinder the establishment of a robust knowledge flow during training, especially in the context of few-shot learning scenarios. These limitations constrain the generalization capability and practical utility of spectroscopic analytical models. To mitigate these issues, we propose an adaptive feedback cross-loop self-KD method to perform preserving spectral information optimization with satisfactory robustness. First, feature space reduction and spectral dataset expansion are automatically achieved by introducing preserving odd-even decomposition while retaining the basic physical and chemical profiles. Then, we introduce a cross-loop architecture with four-teachers to adaptively adjust spectral information during training. The effectiveness of the proposed method is demonstrated through extensive experiments on Infrared (N = 500), Ultraviolet-visible (N = 180), and Raman (N = 2,500) datasets. Our approach significantly mitigates the information degradation caused by spectral processing, thereby alleviating the overfitting issue in spectral few-shot learning. Notably, on the Raman dataset, we achieve a comparable classification accuracy of 96.8% using only 2000 spectra, while the previous result was achieved with a reference dataset of 72 000 spectra. Compared to spectral processing and previous state-of-the-art KD methods, our simple yet powerful method boosts the performance of modern convolutional neural networks across a comprehensive set of metrics, including accuracy, universality, and robustness. The codes are available at https://github.com/wanna-dd/PAD.

Efficient image super-resolution (SR) has witnessed rapid progress thanks to novel lightweight architectures or model compression techniques (e.g., neural architecture search and knowledge distillation). Nevertheless, these methods consume considerable resources or/and neglect to squeeze out the network redundancy at a more fine-grained convolution filter level. Network pruning is a promising alternative to overcome these shortcomings. However, structured pruning is known to be tricky when applied to SR networks because the extensive residual blocks demand the pruned indices of different layers to be the same. Besides, the principled determination of proper layerwise sparsities remains challenging too. In this paper, we present Global Aligned Structured Sparsity Learning (GASSL) to resolve these problems. GASSL has two major components: Hessian-Aided Regularization (HAIR) and Aligned Structured Sparsity Learning (ASSL). HAIR is a regularization-based sparsity auto-selection algorithm with Hessian considered implicitly. A proven proposition is introduced to justify its design. ASSL is for physically pruning SR networks. Particularly, a new penalty term Sparsity Structure Alignment (SSA) is proposed to align the pruned indices of different layers. With GASSL, we design two new efficient single image SR networks of different architecture genres, pushing the efficiency envelope of SR models one step forward. Extensive results demonstrate the merits of GASSL over other recent counterparts.

With the growth of high-dimensional sparse data in web-scale recommender systems, the computational cost to learn high-order feature interaction in CTR prediction task largely increases, which limits the use of high-order interaction models in real industrial applications. Some recent knowledge distillation based methods transfer knowledge from complex teacher models to shallow student models for accelerating the online model inference. However, they suffer from the degradation of model accuracy in knowledge distillation process. It is challenging to balance the efficiency and effectiveness of the shallow student models. To address this problem, we propose a Directed Acyclic Graph Factorization Machine (KD-DAGFM) to learn the high-order feature interactions from existing complex interaction models for CTR prediction via Knowledge Distillation. The proposed lightweight student model DAGFM can learn arbitrary explicit feature interactions from teacher networks, which achieves approximately lossless performance and is proved by a dynamic programming algorithm. Besides, an improved general model KD-DAGFM+ is shown to be effective in distilling both explicit and implicit feature interactions from any complex teacher model. Extensive experiments are conducted on four real-world datasets, including a large-scale industrial dataset from WeChat platform with billions of feature dimensions. KD-DAGFM achieves the best performance with less than 21.5% FLOPs of the state-of-the-art method on both online and offline experiments, showing the superiority of DAGFM to deal with the industrial scale data in CTR prediction task. Our implementation code is available at: https://github.com/RUCAIBox/DAGFM.

Deep learning models have gained prominence in AI research due to their exceptional feature extraction capabilities. In fingerprint-based indoor localization, 1D and 2D convolutional neural networks (CNNs) are frequently utilized for their superior performance and minimal hardware requirements. However, these models primarily extract spatial features and often exhibit degraded real-time performance due to signal inconsistencies caused by shadowing, fading, and multipath effects. To address this, incorporating temporal variations through 3D CNNs can enhance outcomes, albeit with increased computational costs. This study proposes a knowledge distillation (KD) approach using 3D CNNs for indoor localization, to bridge the computational and performance gaps. The proposed KD framework employs a complex teacher model (TM) to train two simpler student models A (SM-A) and B (SM-B). We employ both response-based and feature-based KD methods to optimize the student models’ weights. The proposed techniques are evaluated using the UJILIB dataset collected from two floors of the University of Jaume I library. To enhance model performance, data augmentation is applied, yielding substantial improvements in test accuracy of 10.56% for TM, 16.19% for SM-A, and 16.93% for SM-B. The application of KD results in accuracy improvements of 5.45% for SM-A and 3.35% for SM-B, consequently reducing their average positioning errors (APEs) from 1.070 m and 1.030 m to 0.840 m and 0.944 m, respectively. Compared to KD-based SM-A, TM shows only 3.30% higher positioning accuracy, but with a 65.14% increase in computational complexity. Notably, the KD-enhanced SM-A also achieves a significantly lower APE of 0.84 m compared to the 2D CNN of 1.72 m.

Feedback attention network for cardiac magnetic resonance imaging super-resolution