In the past decade, image restoration has witnessed a remarkable performance leap. However, existing methods still face two critical limitations: i) many methods train distortion-specific models, which struggle to address mixed distortions; ii) most methods employ reconstruction loss during training, with little consideration for the intended use of the restored images (e.g., human viewing, detection, or segmentation), leading to a mismatch between training objectives and application goals. In this paper, we propose a scalable framework for image restoration and enhancement built upon the reinforcement learning (RL) paradigm. At the core of our framework, the RL agent is capable of accurately perceiving distortion contexts and dynamically adjusting diverse restoration operators, which ensures our framework to effectively handle mixed-distortions, and generalize well to new tasks with minor additional cost. Furthermore, RL relaxes the differentiability constraints on optimization objectives, which enables our framework to directly optimize the restoration process toward application goals, thereby bridging the gap between training objectives and downstream demands. Extensive objective and subjective evaluations demonstrate that our method not only generates visually pleasing results for human perception but also significantly improves the performance of machine vision tasks such as detection and segmentation. We further validate the framework’s scalability through generalization evaluations on unseen distortions and multi-objective optimization experiments. The results confirm the model’s superior few-shot and zero-shot capabilities compared to existing methods, as well as its flexibility in addressing multiple competing objectives. Finally, comprehensive ablation studies and hyperparameter analyses provide valuable insights into the design choices and behavior of our framework.

Deep learning-based surface defect segmentation approaches have shown promising performance in recent years. However, segmenting defects with complex shapes, large variations in size, and weakly textured defects with indistinct characteristics still poses significant challenges. In this article, a novel semantic guidance and texture priors based dual-branch surface defect segmentation network (SGTP-Net) is proposed for those issues. Firstly, we construct a feature extraction network combines semantic and texture branches. The semantic branch establishes global contextual relationships, while the texture branch captures local features of defects, this dual-branch ensured the network to extract features from various complex defects. Secondly, we design a feature fusion strategy based on semantic guidance and texture priors. The semantic information is used to guides the output of texture branch. After that, the guided texture information provides valuable edge texture priors for each layer's output in semantic branch. The two branches mutually guide each other for improving ability of weak textures feature extraction. Finally, we run our method on the NEU-Seg, MT-Defect and MSD datasets to conduct a comprehensive comparison with some state-of-the-art general object segmentation models and specialized surface defect segmentation methods. The experimental results show that our SGTP-Net performs well in surface defect detection, offering excellent semantic segmentation accuracy and exhibiting good stability and robustness in detecting various surface defects. Note to Practitioners-This paper was motivated by the problem of surface defect segmentation. Segmentation is foundation of industrial automation testing, surface defect recognition, component size measurement, etc. can all be achieved after segmentation. The existing machine learning based defect segmentation methods rely solely on semantic or texture information for defect representation resulting in poor performance in certain specific scenarios. This paper suggests a new approach combined semantic and texture information. It is not simply a combination of the two, but a dual branch structure that guides each other. Texture information guides the extraction of semantic information, and semantic information re enhances texture information. Finally, those two will be fused for surface defect segmentation. Experiments have shown that our method consistently outperforms other state-of-the-art methods in terms of mean Recall, while also achieving high detection accuracy. It meets the needs of industrial applications. In future research, we will further lightweight the model while maintaining its performance, facilitating its deployment on embedded or mobile devices.

Deep learning has rapidly progressed in synthetic aperture radar (SAR) image object detection. However, complex noise, intricate backgrounds, and multiscale objects impact the performance of SAR image object detection. To solve these problems, in this article, we propose a feature separation and integration network (FSINet) to mitigate noise interference and improve the accuracy of multiscale objects for SAR images. FSINet is constructed in three steps. First, we construct the feature separation module as a preprocessing unit that decouples input images into distinct components, enabling collaboration with the downstream backbone to separate noise and enhance critical information. Subsequently, we propose an adaptive feature integration module designed to fuse cross-layer multiscale features, generating precise feature representations. In addition, a new regression loss called shape normalized Wasserstein distance is used to enhance the detection performance for multiscale objects. To validate the superiority of the proposed FSINet, a series of comparative experiments are conducted on three public SAR image object detection datasets, that is the high-resolution SAR image dataset, SAR-AIRCraft dataset, and MSAR dataset. The experimental results demonstrate the superiority of our method, revealing that our method achieves state-of-the-art performance in SAR image object detection.

Motion estimation in degraded scenes has long been a significant challenge, primarily attributed to substantial scene variations and insufficient training data. Existing approaches typically address this limitation by incorporating additional training strategies or modifying network architectures within conventional frameworks. However, these solutions not only require cumbersome training procedures or additional modal inputs, but also lack generalization capabilities. To address this problem, we propose a unified optical flow estimation framework specifically designed for degraded scenes. In this work, we employ large-scale pre-trained optical flow foundation models as both teacher and student networks. Our objective is to compensate for feature incompleteness during image degradation through pre-trained large models. Subsequently, we leverage supervised signals for fine-tuning and introduce an intra-inter frame distillation method to enable the student network to adapt to diverse cross-domain scenarios. Our proposed methodology provides deeper insights into learning style-invariant features from these learnable fine-tuning layers. Extensive experiments demonstrate that our approach achieves superior generalization performance and state-of-the-art results in degraded scenes (including low-light, rain, fog and other conditions) while requiring minimal training resources.

In unsupervised fusion of hyperspectral and RGB images, the lack of prior knowledge makes it challenging to accurately capture the intrinsic coupling between spatial and spectral degradations, leading to severe spectral distortion and loss of fine spatial details. To solve this problem, we propose a Physics-informed Blind Adaptive Decoupled network (PiBAD) in this letter. PiBAD utilizes a cascaded decoupled initialization strategy to adaptively estimate degradation kernels through independent curvature-regularized spectral response module and total variation-regularized spatial degradation module for HSI and RGB image fusion. A Joint Error Correction mechanism is further introduced to synchronize cross-modal features within a low-rank latent subspace via subspace consistency and gradient constraints. Finally, Progressive Two-Stage Training Strategy ensures the model converges toward a physically realistic solution space. Extensive experiments on Chikusei and Pavia University datasets demonstrate that PiBAD achieves state-of-the-art spatial-spectral fidelity, especially under extreme downsampling scales.

Reconstruction of three-dimensional (3D) scenes from a set of two-dimensional ultrasound images has garnered significant attention in medical research nowadays. Recent studies have demonstrated that neural implicit repre sentation provides a novel and high-fidelity solution. However, most of the existing implicit methods ignore the strong physical correlation between ultrasound reflections and the geometry of human tissues, limiting their abil ity to distinguish between reflections from large tissue interfaces and scattering from microstructures. Moreover, the existing methods do not take into account the scale inconsistency between features of tissues and organs, resulting in loss of detailed information. In this work, we propose a geometry-aware view-dependent implicit 3D ultrasound reconstruction method. First, we design a novel physics-based volume rendering model that can re construct anisotropic and geometry-related reflections more accurately. Then, we employ a multi-resolution hash encoding strategy to capture features at various scales for both tissues and organs, which not only helps differen tiate between broader reflections and fine-grained scattering but also significantly accelerates our reconstruction process. Experiments demonstrate that the proposed method achieves high visual quality and rapid 3D ultrasound reconstruction. Moreover, additional experiments demonstrate the superior performance of the proposed method in distinguishing ultrasound reflections from scattering while preserving the anisotropy of reflections, thereby assisting medical personnel in effectively identifying tissue structures.

Monocular 3D object detection aims to predict object category, position, size, and orientation from a single RGB image. Existing DETR-based monocular 3D detectors suffer from maintaining consistent high-confidence responses due to weak or incomplete target features, resulting in information loss for distant and occluded objects during encoding-decoding. Firstly, to address the core challenge of insufficient target feature perception in complex scenes, we propose a staged iterative monocular 3D detector that progressively refines targets from coarse to fine through multiple paired encoding-decoding stages, significantly improving both feature utilization and network convergence. Furthermore, each stage integrates a dynamic target iteration module that continuously enhances query representation by focusing on high-confidence regional features, thereby enhancing the model's perception of potential targets. Finally, we design a dual-branch depth estimator with parallel global and local processing for a comprehensive representation of the depth feature. Experimental results show that our method achieves superior performance over prior approaches on challenging scenarios (e.g., distant and occluded objects) in the KITTI dataset without auxiliary data, while maintaining competitive accuracy on the nuScenes benchmark under frontal-view settings.

Optical flow estimation has made great progress, but still suffers from motion discontinuities. Although global matching and local regression have made significant efforts, explicit estimation and limited information tend to degrade the performance of optical flow estimation. To alleviate this issue, we propose a novel method named MotionFlow+ for high-accuracy optical flow estimation. First, we propose a cross-appearance enhancement module and an implicit motion prior block, which simultaneously enhance bidirectional features and generate robust motion prior information for the decoding process. Subsequently, we employ the occlusion-aware fusion module during the decoding stage to capture reliable global motion information from motion prior features and cost volume. Next, we align appearance features of the target frame with the reference frame through iteration flow to retrieve missing context crucial for motion decoding. Finally, we introduce an uncertainty-guided optimization module which utilizes uncertainty awareness to identify regions requiring optimization and perform targeted refinement of the iterative flow. Experimental results demonstrate the efficacy of our approach, achieving state-of-the-art performance, particularly outperforming most two-view methods on Sintel Final pass and KITTI-2015 online benchmarks.

Optical flow estimation is a fundamental task in computer vision and visual sensing systems. However, most existing approaches are designed for normal illumination conditions. In low-light scenarios, inherent imaging noise and low contrast lead to noticeable feature degradation and matching ambiguity, which compromises estimation accuracy. To address these issues, this paper proposes SFRC-Flow, a robust low-light optical flow estimation method that integrates spatial–frequency dual-domain feature refinement and motion prior calibration to ensure reliable feature learning and alleviate matching ambiguity. Specifically, the Dual-Domain Refinement Encoder (DDRE) first decomposes shallow spatial features into high-frequency local details and low-frequency global structures. Subsequently, we propose the Global–Local Feature Alignment Module (GLFAM) to resolve spatial and semantic inconsistencies between these decomposed features across different branches via cross-branch feature alignment. Building upon this, we further introduce a cascaded three-stage Frequency-Domain Refinement Module (FDRM) to compensate for the limited long-range modeling capability of spatial-only operations and recover degraded motion features. Finally, we present the Motion Prior-Aware Calibration Module (MPACM) to incorporate motion cues into window-constrained semantic attention. This module produces motion vectors as prior knowledge to calibrate the subsequent flow regression process. Extensive experiments on the FCDN, VBOF, and FLIR ADAS datasets show that SFRC-Flow achieves competitive accuracy for low-light optical flow estimation.

Semantic scene completion networks offer significant potential for reconstructing both geometric and semantic structures of 3D scenes, providing critical environmental understanding for robotic navigation and autonomous driving systems. However, achieving high-precision reconstruction in complex outdoor environments remains challenging. This limitation primarily arises from the difficulty in acquiring sufficiently rich and accurate scene representations solely from sparse sensor data, resulting in a persistent challenge across large-scale spatial domains. To attain the high precision required for robust environmental understanding, we propose leveraging complementary information from multiple modalities. This paper introduces SCG-SSC, a semantic scene completion network that integrates multi-modal priors and employs cross- and self-gated feature control mechanisms. SCG-SSC separates semantics, geometry, and prediction into distinct branches to effectively incorporate these priors. It utilizes two specialized feature control blocks to enhance object recognition while jointly learning semantic and geometric representations. Specifically, the network: (1) Uses pre-trained depth and semantic segmentation networks to provide robust geometric and semantic priors, (2) Incorporates cross-gated blocks in semantic and geometric branches to enhance task-specific feature discrimination beyond initial priors, and (3) Implements self-gated blocks in the fusion branch to refine integrated feature expression through adaptive feature selection, maintaining computational efficiency. Evaluations on the SemanticKITTI and nuScenes-Occupancy benchmarks demonstrate that our approach, empowered by multi-modal priors and gated mechanisms, achieves competitive accuracy against state-of-the-art methods.

With the increasing demand in industrial applications for detecting diverse defect types across multiple product categories, Multi-Class Unsupervised Anomaly Detection (MUAD), which employs a unified model to recognize various anomalies simultaneously, has become a research focus. However, directly applying reconstruction-based methods to multiclass scenarios suffers from the “identical shortcut” problem; single-layer prototype-driven strategies fail to capture both fine-grained textures and high-level semantics. To this end, we propose HIRP-AD (Hierarchical Intrinsic Representative Prototype-Driven Anomaly Detector), a framework that adaptively mines hierarchical intrinsic representative prototypes to guide the reconstruction process. Specifically, HIRP-AD models complementary shallow texture cues and deep semantic cues through hierarchical representative prototypes, thereby guiding the decoder to more precisely reconstruct normal patterns and enabling anomaly detection based on reconstruction discrepancies at both image and pixel levels. Extensive experiments on MVTec-AD, VisA, and Real-IAD demonstrate that HIRP-AD achieves competitive or superior performance compared with existing multi-class anomaly detection methods, validating the effectiveness of hierarchical representative prototype mining for unified anomaly detection and localization. In addition, by replacing LayerNorm with DynamicTanh (DyT), HIRP-AD reduces normalization-related computation and further improves inference efficiency, providing a hardware-friendly solution for MUAD.

BACKGROUND:Accurate whole-brain tissue segmentation from MRI is fundamental in medical image analysis. Although existing 3D segmentation methods perform well on thin-slice MRI data, their performance often decreases on thick-slice images due to low through-plane resolution and missing inter-slice anatomical information. NEW METHOD:We developed a bi-directional super-resolution reconstruction (BSR)-assisted 3D segmentation framework for thick-slice brain MRI. First, a general whole-brain segmentation model based on nnU-Net was trained using thin-slice MRI data to obtain robust cross-dataset segmentation capability. Then, the proposed BSR network was used to recover missing inter-slice information from thick-slice MRI and generate thin-slice-like 3D images with improved anatomical continuity. The reconstructed high-quality slices were inserted into the inter-slice gaps of thick-slice MRI data as supplementary anatomical information for subsequent segmentation. RESULTS:The proposed framework was evaluated on multiple public brain MRI datasets, including IBSR18, LPBA40, and OASIS. Experimental results demonstrated that our method effectively improved whole-brain segmentation performance on thick-slice MRI data. When LPBA40 was used as the training dataset, our framework achieved a 3.6 % improvement over the original nnU-Net on the challenging IBSR18 thick-slice dataset. On another thick-slice dataset, OASIS-3, the proposed method obtained a 1.2 % accuracy improvement. COMPARISON WITH EXISTING METHODS:Compared with conventional nnU-Net-based segmentation, our framework improves robustness on thick-slice MRI by explicitly compensating for missing inter-slice anatomical information before segmentation. CONCLUSION:The proposed BSR-assisted framework provides an effective solution for whole-brain segmentation from thick-slice MRI and may facilitate more reliable brain tissue analysis in clinical scenarios.

Monocular depth estimation is widely used for navigation and scene understanding, yet deployment requires balancing accuracy with predictable runtime and compact models. Many recent lightweight designs pair depthwise separable convolutions with transformer components to boost accuracy, which typically introduces a more diverse operator set and can make realized throughput more dependent on the deployment backend in practice. Instead, we revisit recurrent refinement from a deployment-oriented perspective and introduce R-TAFM, purely convolutional framework that performs iterative depth refinement at a fixed working resolution with a parameter-shared decoder. We further derive a deployment-mode variant, R-TAFM-Fast, which is trained with recurrent supervision yet reduces inference to a single decoder pass, lowering latency on commodity GPU and on a Jetson-class embedded GPU. For self-supervised learning, we introduce an adaptive reprojection objective that jointly handles occlusions and independently moving objects without auxiliary tasks, and a neighborhood-consistent correction of auto-masked stationary pixels to prevent supervision collapse in homogeneous regions. Both quantitative benchmarks and qualitative assessments demonstrate that, with ≈ 3.2M parameters, R-TAFM achieves accuracy comparable to or exceeding recent lightweight state-of-the-art methods, using only standard convolutions.

Multi-source image fusion combines infrared and visible information to improve scene perception in applications such as drone reconnaissance and autonomous driving. However, most existing infrared-visible image fusion methods are developed under ideal imaging assumptions. In adverse environments, visible images often lose structural and textural details, whereas infrared images are affected by noise, stripe artifacts, and low contrast, leading to degraded fusion quality and weakened downstream perception performance. To address these limitations, we propose a unified Degradation-aware Restoration and Detail-preserving Fusion Network (DRDFNet), which consists of a Degradation-Aware Restoration Transformer and a Detail-Preserving Fusion Mamba. The restoration branch uses a Compound Degradation Restoration Module (CDRM) to remove complex degradations, while the fusion branch employs a Dynamic Feature Fusion Module (DFFM) to integrate local complementary cues and global correlations across modalities. A two-stage training strategy is further introduced to reduce the optimization conflict between restoration and fusion. In addition, we construct DIVIF, a large-scale degraded IVIF benchmark generated by a physics-based imaging simulator. Experiments on the DIVIF and AWMM-100k benchmarks demonstrate that DRDFNet achieves robust and competitive performance compared with SOTA methods. Both the dataset and source code will be made publicly available at https://github.com/Liupeng97/DRDFNet.

Instance segmentation plays a vital role in practical applications, including autonomous driving and robotic vision. Among deep learning methods, query-based approaches have recently attracted significant attention due to their effectiveness in building end-to-end pipelines that eliminate complex post-processing steps. However, existing methods often fail to strike an optimal balance between accuracy and real-time performance. To address this challenge, we propose SAP-DQR, a framework integrating Spatial-Adaptive Pyramid features and Adaptive Query Reorganization to optimize the speed-accuracy trade-off. First, we design a Spatial-Adaptive Pyramid Feature Network (SAP-FPN) that integrates Asymmetric and Separable Multi-scale Pooling (ASMP) and Spatial-Channel Grouped Fusion (SCG). This design significantly accelerates and improves the efficiency of multi-scale feature processing. Additionally, we introduce a novel lightweight Global Semantic-guided Multi-stage Adaptive Query Reorganization (GSMR) mechanism that effectively provides finer-grained, higher-quality adaptive queries for various decoder stages. Our method is the first ResNet-50-based approach to surpass 40 AP on COCO while achieving real-time inference ( ≥ 30 FPS), delivering an optimal speed-accuracy trade-off for latency-sensitive applications like autonomous driving. It significantly outperforms current mainstream methods, setting a new state-of-the-art for real-time instance segmentation with optimized speed-accuracy trade-off.

Industrial Anomaly Detection (IAD) is pivotal in smart manufacturing, and recent prototype-based reconstruction paradigms (e.g., INP-Former) demonstrate strong performance in a generalized unsupervised setting. However, a fixed number of prototypes cannot balance coverage of normal patterns in complex scenes with computational efficiency in simple ones; meanwhile, relying solely on pixel-level reconstruction error is not sufficiently sensitive to structural disruptions. Therefore, we propose DNP-Former, introducing two components: (i) Dynamic-M with Dual-Stage Sparsity (DMDS), which imposes soft sparsity on prototype importance during training and performs threshold-based hard pruning at inference, thereby enabling sample-wise adaptive determination of the effective prototype count M′; and (ii) a Cross-Scale Structural Coherence Loss (CSCL) that applies cross-scale structural consistency regularization to the reconstruction, suppressing illumination artifacts and highlighting genuine structural degradations. On MVTec-AD and VisA, DNP-Former matches or slightly exceeds image-level performance and significantly improves pixel-level metrics over SOTA, while reducing FLOPs by 37.5% and inference latency by 24.4%. The paradigm couples sample-wise adaptive computation with interpretability, delivering a favorable efficiency–accuracy trade-off and facilitating deployment in generalized IAD.

Optical flow, as a feature representation of optical patterns, has been extensively studied and applied in motion detection. However, its application in defect detection remains rarely explored. To address the reliance of machine vision-based automated defect detection on large labelled datasets, which poses significant challenges on the detection executed to specialised components or online inspection scenes, we propose an alternative way to apply optical flow estimation as a method to discriminate the varying rates of mode clusters. Assuming optical flow components are values of the probability density function of optical mode, we consider optical flow estimation transitions through which optical mode characteristics are conveyed. Thereby, the modulation theory of stationary random processes (SRPs) is introduced to describe the basic equation of optical flow estimation. We remodel the basic equation according to the communication principle and shorten the calculation time. To better utilise the limited bandwidth of pattern space and achieve accurate pattern feature recognition within it, we consider optical flow estimation as an image-based communication system and find the correspondences between them. By doing so, quantisation schemes and pyramid construction are evaluated under convex optimisation of optical flow estimation, and experimental results confirm quantisation scheme is feasible of the two. A modified ideal quantisation scheme is suggested to perform robust optical flow estimation. Experimental results reveal quantisation provided by Charbonnier compresses the output bandwidth in a wide range of parameter settings and guarantees an optical flow method with higher accuracy for automated machine vision detection executed in an unprepared industrial scene.

Monocular 3D object detection offers significant potential for autonomous systems due to its inherent cost-effectiveness and scalability. While DETR-based architectures excel in 2D vision tasks, critical limitations persist in extending them effectively to monocular 3D detection, as evidenced in existing frameworks like MonoDETR and MonoDGP. These methods typically suffer from inefficient serial fusion of multimodal features and lack iterative refinement mechanisms, limiting their performance, especially for mid-to-long range targets. To overcome these shortcomings, we propose Iter3DDet, a novel depth-guided iterative refinement framework that integrates fine-grained feature fusion to significantly enhance detection performance. The core novelty of our approach lies in two key innovations: (1) A hybrid feature encoder combining MonoDGP’s region segmentation head with MonoDETR’s visual backbone, augmented by a multi-scale context attention module that dynamically aggregates structural and semantic cues across pyramid levels, eliminating heuristic fusion rules; (2) A depth-guided adaptive cross-modal decoder that iteratively fuses depth and context features through prioritized attention mechanisms, coupled with a novel iterative refinement training strategy that progressively refines 3D detection hypotheses, substantially improving accuracy across targets of varying difficulty levels. Extensive experiments on the KITTI, nuScenes, and Waymo benchmarks demonstrate Iter3DDet’s state-of-the-art performance, validating the effectiveness of our iterative refinement paradigm. The code will be open-sourced at https://github.com/PCwenyue.

Most existing emotion recognition studies rely on discrete stimulus paradigms under controlled laboratory conditions, failing to capture the inherently dynamic and transitional nature of emotional states encountered in real-world operational environments. To address this gap, we propose a novel approach for emotion transition recognition using multimodal physiological signal fusion. Thirty participants, including 10 licensed pilots and 20 non-pilots controls, were recruited to view 48 video stimuli de-signed to elicit six distinct emotional transitions. We introduce a two-stage progressive recognition framework to enhance performance in emotional transition modeling. In Stage I, differential entropy features extracted from EEG signals are integrated with Big Five personality traits to construct an occupation-based trait classification model, capturing cognitive-emotional patterns associated with occupational background and providing individualized priors. In Stage II, we develop the temporal spatial embed ding bidirectional cross-attention fusion network (TSE-BCAF), which employs parallel convolutional modules to extract deep spatial features from multimodal EEG and EOG signals. These features are fused via a bidirectional cross-attention mechanism to model inter-modal interactions, then processed by a Trans former-based temporal encoder to capture the evolving dynamics of emotional transitions. Experimental results demonstrate that the proposed framework significantly outperforms unimodal and state-of-the-art multimodal baselines in both accuracy and robustness. The findings highlight the potential of our approach for real-world applications in aviation psychology and other high stakes domains requiring adaptive emotional regulation.

Semi-supervised learning (SSL) is an effective solution for medical image segmentation with limited annotations. Existing SSL methods mainly rely on pseudo-labels generated by teacher-student supervision or cross-network consistency. However, these methods lack an explicit structural reference for judging pseudo-label quality. Low-quality pseudo-labels may lead to unreliable training, error accumulation and confirmation bias when processing unlabeled data with substantial appearance variations. To address this issue, we proposed OFD-Net, a teacher-free single-network framework for reliable semi-supervised medical image segmentation. OFD-Net employs an Orthogonal Feature Disentanglement Module (OFDM) to capture OFD features for reliable SSL by disentangling unlabeled data into background and foreground representations with a reliable structural distribution, thereby effectively reducing error accumulation and alleviating confirmation bias among unlabeled data. Specifically, OFD-Net explicitly employs a Disentanglement Guidance Module (DGM) to inject the resulting structural priors of foreground-background into the decoder by deformable convolution processing, and outputs predictions with clearer foreground representations. Based on DGM and the OFDM, we further develop a reliability-aware pseudo-label learning mechanism that evaluates unlabeled supervision according to the structural consistency between the main prediction and the disentangled foreground-background responses, and then down-weights unreliable regions during training. Extensive experiments on four public medical image segmentation benchmarks, namely ISIC-2016, Kvasir-SEG, Synapse, and ACDC, validate the effectiveness of OFD-Net. These results confirm that orthogonal foreground-background disentanglement enables OFD-Net to establish an efficient and reliable training paradigm within a teacher-free single-network framework.