X-ray illicit object detection using hybrid CNN-transformer neural network architectures

In the field of X-ray security applications, even the smallest details can significantly impact outcomes. Objects that are heavily occluded or intentionally concealed pose a great challenge for detection, whether by human observation or through advanced technological applications. While certain Deep Learning (DL) architectures demonstrate strong performance in processing local information, such as Convolutional Neural Networks (CNNs), others excel in handling distant information, e.g., transformers. In X-ray security imaging the literature has been dominated by the use of CNN-based methods, while the integration of the two aforementioned leading architectures has not been sufficiently explored. In this paper, various hybrid CNN-transformer architectures are evaluated against a common CNN object detection baseline, namely YOLOv8. In particular, a CNN (HGNetV2) and a hybrid CNN-transformer (Next-ViT-S) backbone are combined with different CNN/transformer detection heads (YOLOv8 and RT-DETR). The resulting architectures are comparatively evaluated on three challenging public X-ray inspection datasets, namely EDS, HiXray, and PIDray. Interestingly, while the YOLOv8 detector with its default backbone (CSP-DarkNet53) is generally shown to be advantageous on the HiXray and PIDray datasets, when a domain distribution shift is incorporated in the X-ray images (as happens in the EDS datasets), hybrid CNN-transformer architectures exhibit increased robustness. Detailed comparative evaluation results, including object-level detection performance and object-size error analysis, demonstrate the strengths and weaknesses of each architectural combination and suggest guidelines for future research. The source code and network weights of the models employed in this study are available at https://github.com/jgenc/xray-comparative-evaluation.

Can We Achieve Efficient Diffusion without Self-Attention? Distilling Self-Attention into Convolutions

Contemporary diffusion models built upon U-Net or Diffusion Transformer (DiT) architectures have revolutionized image generation through transformer-based attention mechanisms. The prevailing paradigm has commonly employed self-attention with quadratic computational complexity to handle global spatial relationships in complex images, thereby synthesizing high-fidelity images with coherent visual semantics.Contrary to conventional wisdom, our systematic layer-wise analysis reveals an interesting discrepancy: self-attention in pre-trained diffusion models predominantly exhibits localized attention patterns, closely resembling convolutional inductive biases. This suggests that global interactions in self-attention may be less critical than commonly assumed.Driven by this, we propose \(\Delta\)ConvFusion to replace conventional self-attention modules with Pyramid Convolution Blocks (\(\Delta\)ConvBlocks).By distilling attention patterns into localized convolutional operations while keeping other components frozen, \(\Delta\)ConvFusion achieves performance comparable to transformer-based counterparts while reducing computational cost by 6929$\times$ and surpassing LinFusion by 5.42$\times$ in efficiency--all without compromising generative fidelity.

Rethinking Visual Layer Selection in Multimodal LLMs

Multimodal large language models (MLLMs) have achieved impressive performance across a wide range of tasks, typically using CLIP-ViT as their visual encoder due to its strong text-image alignment capabilities. While prior studies suggest that different CLIP-ViT layers capture different types of information, with shallower layers focusing on fine visual details and deeper layers aligning more closely with textual semantics, most MLLMs still select visual features based on empirical heuristics rather than systematic analysis. In this work, we propose a Layer-wise Representation Similarity approach to group CLIP-ViT layers with similar behaviors into {shallow, middle, and deep} categories and assess their impact on MLLM performance. Building on this foundation, we revisit the visual layer selection problem in MLLMs at scale, training LLaVA-style models ranging from 1.4B to 7B parameters. Through extensive experiments across 10 datasets and 4 tasks, we find that: (1) deep layers are essential for OCR tasks; (2) shallow and middle layers substantially outperform deep layers on reasoning tasks involving counting, positioning, and object localization; (3) a lightweight fusion of features across shallow, middle, and deep layers consistently outperforms specialized fusion baselines and single-layer selections, achieving gains on 9 out of 10 datasets. Our work offers the first principled study of visual layer selection in MLLMs, laying the groundwork for deeper investigations into visual representation learning for MLLMs.

Image deidentification in the XNAT ecosystem: use cases and solutions

XNAT is a server-based data management platform widely used in academia for curating large databases of DICOM images for research projects. We describe in detail a deidentification workflow for DICOM data using facilities in XNAT, together with independent tools in the XNAT "ecosystem". We list different contexts in which deidentification might be needed, based on our prior experience. The starting point for participation in the Medical Image De-Identification Benchmark (MIDI-B) challenge was a set of pre-existing local methodologies, which were adapted during the validation phase of the challenge. Our result in the test phase was 97.91\%, considerably lower than our peers, due largely to an arcane technical incompatibility of our methodology with the challenge's Synapse platform, which prevented us receiving feedback during the validation phase. Post-submission, additional discrepancy reports from the organisers and via the MIDI-B Continuous Benchmarking facility, enabled us to improve this score significantly to 99.61\%. An entirely rule-based approach was shown to be capable of removing all name-related information in the test corpus, but exhibited failures in dealing fully with address data. Initial experiments using published machine-learning models to remove addresses were partially successful but showed the models to be "over-aggressive" on other types of free-text data, leading to a slight overall degradation in performance to 99.54\%. Future development will therefore focus on improving address-recognition capabilities, but also on better removal of identifiable data burned into the image pixels. Several technical aspects relating to the "answer key" are still under discussion with the challenge organisers, but we estimate that our percentage of genuine deidentification failures on the MIDI-B test corpus currently stands at 0.19\%. (Abridged from original for arXiv submission)

DiffusionRIR: Room Impulse Response Interpolation using Diffusion Models

Room Impulse Responses (RIRs) characterize acoustic environments and are crucial in multiple audio signal processing tasks. High-quality RIR estimates drive applications such as virtual microphones, sound source localization, augmented reality, and data augmentation. However, obtaining RIR measurements with high spatial resolution is resource-intensive, making it impractical for large spaces or when dense sampling is required. This research addresses the challenge of estimating RIRs at unmeasured locations within a room using Denoising Diffusion Probabilistic Models (DDPM). Our method leverages the analogy between RIR matrices and image inpainting, transforming RIR data into a format suitable for diffusion-based reconstruction. Using simulated RIR data based on the image method, we demonstrate our approach's effectiveness on microphone arrays of different curvatures, from linear to semi-circular. Our method successfully reconstructs missing RIRs, even in large gaps between microphones. Under these conditions, it achieves accurate reconstruction, significantly outperforming baseline Spline Cubic Interpolation in terms of Normalized Mean Square Error and Cosine Distance between actual and interpolated RIRs. This research highlights the potential of using generative models for effective RIR interpolation, paving the way for generating additional data from limited real-world measurements.

Reinforcement Learning-Based Heterogeneous Multi-Task Optimization in Semantic Broadcast Communications

Semantic broadcast communications (Semantic BC) for image transmission have achieved significant performance gains for single-task scenarios. Nevertheless, extending these methods to multi-task scenarios remains challenging, as different tasks typically require distinct objective functions, leading to potential conflicts within the shared encoder. In this paper, we propose a tri-level reinforcement learning (RL)-based multi-task Semantic BC framework, termed SemanticBC-TriRL, which effectively resolves such conflicts and enables the simultaneous support of multiple downstream tasks at the receiver side, including image classification and content reconstruction tasks. Specifically, the proposed framework employs a bottom-up tri-level alternating learning strategy, formulated as a constrained multi-objective optimization problem. At the first level, task-specific decoders are locally optimized using supervised learning. At the second level, the shared encoder is updated via proximal policy optimization (PPO), guided by task-oriented rewards. At the third level, a multi-gradient aggregation-based task weighting module adaptively adjusts task priorities and steers the encoder optimization. Through this hierarchical learning process, the encoder and decoders are alternately trained, and the three levels are cohesively integrated via constrained learning objective. Besides, the convergence of SemanticBC-TriRL is also theoretically established. Extensive simulation results demonstrate the superior performance of the proposed framework across diverse channel conditions, particularly in low SNR regimes, and confirm its scalability with increasing numbers of receivers.

Curiosity Driven Exploration to Optimize Structure-Property Learning in Microscopy

Rapidly determining structure-property correlations in materials is an important challenge in better understanding fundamental mechanisms and greatly assists in materials design. In microscopy, imaging data provides a direct measurement of the local structure, while spectroscopic measurements provide relevant functional property information. Deep kernel active learning approaches have been utilized to rapidly map local structure to functional properties in microscopy experiments, but are computationally expensive for multi-dimensional and correlated output spaces. Here, we present an alternative lightweight curiosity algorithm which actively samples regions with unexplored structure-property relations, utilizing a deep-learning based surrogate model for error prediction. We show that the algorithm outperforms random sampling for predicting properties from structures, and provides a convenient tool for efficient mapping of structure-property relationships in materials science.

A Transformer-based Multimodal Fusion Model for Efficient Crowd Counting Using Visual and Wireless Signals

Current crowd-counting models often rely on single-modal inputs, such as visual images or wireless signal data, which can result in significant information loss and suboptimal recognition performance. To address these shortcomings, we propose TransFusion, a novel multimodal fusion-based crowd-counting model that integrates Channel State Information (CSI) with image data. By leveraging the powerful capabilities of Transformer networks, TransFusion effectively combines these two distinct data modalities, enabling the capture of comprehensive global contextual information that is critical for accurate crowd estimation. However, while transformers are well capable of capturing global features, they potentially fail to identify finer-grained, local details essential for precise crowd counting. To mitigate this, we incorporate Convolutional Neural Networks (CNNs) into the model architecture, enhancing its ability to extract detailed local features that complement the global context provided by the Transformer. Extensive experimental evaluations demonstrate that TransFusion achieves high accuracy with minimal counting errors while maintaining superior efficiency.

Magnifier: A Multi-grained Neural Network-based Architecture for Burned Area Delineation

In crisis management and remote sensing, image segmentation plays a crucial role, enabling tasks like disaster response and emergency planning by analyzing visual data. Neural networks are able to analyze satellite acquisitions and determine which areas were affected by a catastrophic event. The problem in their development in this context is the data scarcity and the lack of extensive benchmark datasets, limiting the capabilities of training large neural network models. In this paper, we propose a novel methodology, namely Magnifier, to improve segmentation performance with limited data availability. The Magnifier methodology is applicable to any existing encoder-decoder architecture, as it extends a model by merging information at different contextual levels through a dual-encoder approach: a local and global encoder. Magnifier analyzes the input data twice using the dual-encoder approach. In particular, the local and global encoders extract information from the same input at different granularities. This allows Magnifier to extract more information than the other approaches given the same set of input images. Magnifier improves the quality of the results of +2.65% on average IoU while leading to a restrained increase in terms of the number of trainable parameters compared to the original model. We evaluated our proposed approach with state-of-the-art burned area segmentation models, demonstrating, on average, comparable or better performances in less than half of the GFLOPs.

LR-IAD:Mask-Free Industrial Anomaly Detection with Logical Reasoning

Industrial Anomaly Detection (IAD) is critical for ensuring product quality by identifying defects. Traditional methods such as feature embedding and reconstruction-based approaches require large datasets and struggle with scalability. Existing vision-language models (VLMs) and Multimodal Large Language Models (MLLMs) address some limitations but rely on mask annotations, leading to high implementation costs and false positives. Additionally, industrial datasets like MVTec-AD and VisA suffer from severe class imbalance, with defect samples constituting only 23.8% and 11.1% of total data respectively. To address these challenges, we propose a reward function that dynamically prioritizes rare defect patterns during training to handle class imbalance. We also introduce a mask-free reasoning framework using Chain of Thought (CoT) and Group Relative Policy Optimization (GRPO) mechanisms, enabling anomaly detection directly from raw images without annotated masks. This approach generates interpretable step-by-step explanations for defect localization. Our method achieves state-of-the-art performance, outperforming prior approaches by 36% in accuracy on MVTec-AD and 16% on VisA. By eliminating mask dependency and reducing costs while providing explainable outputs, this work advances industrial anomaly detection and supports scalable quality control in manufacturing. Code to reproduce the experiment is available at https://github.com/LilaKen/LR-IAD.