In this paper, a novel satellite image segmentation technique based on dynamic Harris hawks optimization with a mutation mechanism (DHHO/M) is proposed. Compared with the original Harris hawks optimization (HHO), the dynamic control parameter strategy and mutation operator used in DHHO/M can avoid falling into the local optimum and efficiently enhance the search capability. To evaluate the performance of the proposed method, a series of experiments are carried out on various satellite images. Eight advanced thresholding approaches are selected for comparison. Three criteria are adopted to determine the segmentation thresholds, namely Kapur’s entropy, Tsallis entropy, and Otsu between-class variance. Furthermore, four oil pollution images are used to further assess the practicality and feasibility of the proposed method on real engineering problem. The experimental results illustrate that the DHHO/M based thresholding technique is superior to others in the following three aspects: fitness function evaluation, image segmentation effect, and statistical tests.

An efficient satellite image segmentation method based on a hybrid grasshopper optimization algorithm (GOA) and minimum cross entropy (MCE) is proposed in this paper. The proposal is known as GOA–jDE, and it merges GOA with self-adaptive differential evolution (jDE) to improve the search efficiency, preserving the population diversity especially in the later iterations. A series of experiments is conducted on various satellite images for evaluating the performance of the algorithm. Both low and high levels of the segmentation are taken into account, increasing the dimensionality of the problem. The proposed approach is compared with the standard color image thresholding methods, as well as the advanced satellite image thresholding techniques based on different criteria. Friedman test and Wilcoxon’s rank sum test are performed to assess the significant difference between the algorithms. The superiority of the proposed method is illustrated from different aspects, such as average fitness function value, peak signal to noise ratio (PSNR), structural similarity index (SSIM), feature similarity index (FSIM), standard deviation (STD), convergence performance, and computation time. Furthermore, natural images from the Berkeley segmentation dataset are also used to validate the strong robustness of the proposed method.

A novel multilevel threshold segmentation method for color satellite images based on Masi entropy is proposed in this paper. Lévy multiverse optimization algorithm (LMVO) has a strong advantage over the traditional multiverse optimization algorithm (MVO) in finding the optimal solution for the segmentation in the three channels of an RGB image. As the work advancement introduces a Lévy multiverse optimization algorithm which uses tournament selection instead of roulette wheel selection, and updates some formulas in the algorithm with mutation factor. Then, the proposal is called TLMVO, and another advantage is that the population diversity of the algorithm in the latest iterations is maintained. The Masi entropy is used as an application and combined with the improved TLMVO algorithm for satellite color image segmentation. Masi entropy combines the additivity of Renyi entropy and the non-extensibility of Tsallis entropy. By increasing the number of thesholds, the quality of segmenttion becomes better, then the dimensionality of the problem also increases. Fitness function value, average CPU running time, Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM) and Feature Similarity Index (FSIM) were used to evaluate the segmentation results. Further statistical evaluation was given by Wilcoxon’s rank sum test and Friedman test. The experimental results show that the TLMVO algorithm has wide adaptability to high-dimensional optimization problems, and has obvious advantages in objective function value, image quality detection, convergence performance and robustness.

Traffic sign classification (TSC) based on deep neural networks (DNNs) plays a crucial role in the perception subsystem of autonomous driving systems (ADSs). However, studies reveal that the TSC system can make dangerous and potentially fatal errors under adversarial attacks. Existing defense strategies, such as adversarial training (AT), have demonstrated effectiveness but struggle to generalize across diverse attack scenarios. Recent advancements in self-supervised learning (SSL), particularly adversarial contrastive learning (ACL) methods, have demonstrated strong potential in enhancing robustness and generalization compared to AT. However, conventional ACL methods lack mechanisms to ensure effective defense transferability across different learning stages. To address this, we propose a robust pre-training–fine-tuning algorithm for meta-adversarial defense (RPF-MAD), designed to enhance the sustainability of adversarial robustness throughout the learning pipeline. Dual-track meta-adversarial pre-training (Dual-MAP) integrates meta-learning with ACL methods, which improves the generalization ability of the upstream model to different adversarial conditions. Meanwhile, adaptive variance anchoring robust fine-tuning (AVA-RFT) utilizes adaptive prototype variance regularization to stabilize feature representations and reinforce the generalizable defense capabilities of the downstream model. Leveraging the meta-adversarial defense benchmark (MAD) dataset, RPF-MAD ensures comprehensive robustness against multiple attack types. Extensive experiments across eight ACL methods and three robust fine-tuning (RFT) techniques demonstrate that RPF-MAD significantly improves both standard accuracy ($SA$) by 1.53% and robust accuracy ($RA$) by 2.64%, effectively enhances the lifelong adversarial resilience of TSC models, achieves a 13.77% improvement in the equilibrium defense success rate ($EDSR$), and reduces the attack success rate ($ASR$) by 9.74%, outperforming state-of-the-art (SOTA) defense methods.

Vision Language Models (VLMs) have advanced perception in autonomous driving (AD), but they remain vulnerable to adversarial threats. These risks range from localized physical patches to imperceptible global perturbations. Existing defense methods for VLMs remain limited and often fail to reconcile robustness with clean-sample performance. To bridge these gaps, we propose NutVLM, a comprehensive self-adaptive defense framework designed to secure the entire perception-decision lifecycle. Specifically, we first employ NutNet++ as a sentinel, which is a unified detection-purification mechanism. It identifies benign samples, local patches, and global perturbations through three-way classification. Subsequently, localized threats are purified via efficient grayscale masking, while global perturbations trigger Expert-guided Adversarial Prompt Tuning (EAPT). Instead of the costly parameter updates of full-model fine-tuning, EAPT generates "corrective driving prompts" via gradient-based latent optimization and discrete projection. These prompts refocus the VLM’s attention without requiring exhaustive full-model retraining. Evaluated on the Dolphins benchmark, our NutVLM yields a 4.89% improvement in overall metrics (e.g., Accuracy, Language Score, and GPT Score). These results validate NutVLM as a scalable security solution for intelligent transportation. Our code is available at https://github.com/PXX/NutVLM.

Surface defect detection is a crucial method for controlling product quality, and convolutional neural networks(CNN) have shown great potential in this field. However, the researchers found that adversarial examples can cause the CNN to produce inaccurate results. It remains to be studied whether CNN can be applied in important industries. In order to improve the adversarial defense capability of neural networks, this paper proposes a neural network training framework based on cross-data space adversarial defense(CDSAD) for classified neural networks. The framework CDSAD trains a network that can adapt to multiple data distributions. During the inference process, the network employs multi-perspective evaluation methods to defend, warn and detect adversarial examples. The defensive capability of the CDSAD was tested against multiple adversarial attack algorithms using the CIFAR 10 and triode surface defect datasets. Experiment results demonstrate that the CDSAD defends against multiple adversarial attack algorithms without requiring any information about the opponent. The defensive effect is further improved when the framework acquires opponent information for adversarial identification module(AIM) training.