AI-driven disease diagnosis relies on large datasets and advanced algorithms, but data scarcity poses challenges for infectious disease detection. To address this, we propose the direct-self-attention Wasserstein generative adversarial network (DSAWGAN), which enhances diagnostic accuracy with limited data. By integrating attention mechanisms and Wasserstein distance optimization, DSAWGAN improves convergence speed, stability, and image quality. We compared it with traditional data augmentation methods using classification neural networks. The trained models were then deployed in a mobile app for rapid, portable, and cost-effective diagnostics. With only half the original data (n = 1500), accuracy increased from 98.00% to 99.33%. Even with just 10% (n = 300), the model maintained 92.67% accuracy, demonstrating its effectiveness in resource-limited settings.
On-chip hyperspectral imaging (HSI) captures substantially richer spectral signatures than conventional RGB sensors, thereby enhancing semantic segmentation performance for autonomous driving and intelligent transportation systems. However, the elevated dimensionality of HSI data poses considerable computational burdens, necessitating effective compression. We address this challenge by introducing MobileSACNet, a MobileUNet-based framework equipped with a Spectral Attention Compressor (SAC). This nonlinear compression module integrates multi-stage 3D convolutions and dynamic channel attention through two complementary branches: (1) a spectral compression branch employing depthwise separable 3D layers with squeeze-excitation attention to refine spectral representations, and (2) a residual pathway that preserves spatial-spectral correlations via identity mappings to mitigate information loss. Evaluations on an on-chip HSI automotive dataset reveal that MobileSACNet substantially surpasses state-of-the-art baselines, especially in complex road scenarios. Moreover, ablation studies illuminate the trade-off between computational efficiency and segmentation fidelity, offering critical insights for real-world deployment in intelligent transport and autonomous vehicle surveillance.
This paper proposes a binary classification framework for chronic endometritis diagnosis in hysteroscopy images. While hysteroscopy is less invasive than a histopathological examination, appearance feature-based diagnosis is not easy due to very similar appearance features of positive and negative samples observed in low-quality images. This work explores various components in our framework to explore the possibility of a less invasive diagnosis as a medical application of machine vision. Our experiments demonstrate that the classification performance of our framework is comparable to manual diagnosis using known symptoms in terms of several metrics.
Recognition of fine-grained actions involving similar tools and motions in egocentric videos is a challenging task in computer vision. In general, to distinguish such actions, humans often focus on the resultant state of the manipulated objects. From this perspective, we propose a novel fine-grained action recognition method that leverages object states in egocentric cooking videos through three key components: 1) precise extraction of the region containing only the food ingredient, reducing background noise; 2) selection of key frames ("End State frames") that clearly reflect the post-action state of the object; and 3) augmentation of training data using generated images. In this paper, we empirically demonstrate the effectiveness of the proposed method with our own dataset.
Laser Speckle Authentication (LSA) has been widely studied, yet the impact of the optical system size on its robustness against object misalignment remains underexplored. To clarify this, we evaluated the effects of perpendicular and parallel misalignments using six optical systems combining lens-based and lensless configurations. Our findings reveal that lens-based systems exhibit superior robustness against misalignment although they require a larger system size. This research provides insights for developing compact, robust authentication systems under practical size constraints.
Mainstream image manipulation-based data augmentation methods undermine the integrity of extracted features, which limits their effectiveness for pixel-level image restoration tasks. In this paper, a reflection-specific data augmentation method, called IRR-RADA, for image reflection removal is proposed. Within it, the Reflection-Aware Saliency Map (RASM) accurately localizes reflection-affected regions by leveraging differential saliency effects, while the Adaptive Curriculum Learning (ACL) strategy gradually adjusts augmentation intensity based on reflection complexity and training stage. By integrating these two operations, the model is compelled to learn precise reflection removal while preserving the inherent structure of the target image. Experiments demonstrate that IRR-RADA achieves an average PSNR boost of 0.91dB compared to the baseline model without data augmentation.
This research proposes a low-latency real-time audio-driven talking head generation (THG) method based on future speech feature prediction. Most traditional methods often suffer from high latency, hindering real-time interaction because they inevitably require future audio context to maintain generation quality. To address this, we predict future speech frames from mel spectrogram features, enabling the generation of facial animations in advance and reducing latency. Our approach integrates the Neural Radiance Field (NeRF) model for high-fidelity facial animation, ensuring precise synchronization with speech input. Extensive experiments demonstrate the high prediction accuracy of the proposed future speech feature prediction module. Furthermore, comparisons with previous works confirm that our method significantly reduces latency while improving video quality. The key contribution of this work is the introduction of future speech feature prediction to real-time THG, effectively reducing synchronization latency while maintaining high video quality.
Point Goal Navigation (PGN) has advanced significantly in robot vision. However, current methods are designed for single-robot systems and rely on first-person view (FPV) state-action representations, making generalization difficult across robot architectures. We propose a universal state-action representation based on third-person view (TPV) local maps to enable teacher-student knowledge transfer from a mapper (teacher) to a map-user robot (student). Our method is validated in worst-case yet practical scenarios, where heterogeneous robots with different platforms (e.g., differentiable vs. non-differentiable models, C++ vs. Python) must collaborate—an understudied setting in prior work.
The use of large-scale models such as neural networks has been increasing in recent years. When using NN, it is generally necessary to prepare a large amount of data. However, collecting a large number of images for plants poses a challenge. Therefore, we demonstrate the effectiveness of combining a geometric feature with NN, even with a small dataset. This paper proposes a method for predicting the age of komatsuna leaves using a combination of a Neural Network model and Hu moments that have traditionally been used as geometric features in computer vision. We used ViT and ResNet as Neural Network models. The results show that Hu moments with NN consistently achieve higher accuracy than only NN across different dataset sizes. Specifically, Hu moments with ViT attained 96.3% accuracy, exceeding the 84.2% of the ResNet-based model relying solely on image features. As shown in Figure 1, the advantage of incorporating Hu moments with NN remains evident regardless of the dataset size.
We propose a machine learning-based approach with feature selection for automatic finger tapping rating in the Movement Disorder Society-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) Part III assessment. A total of 160 video clips from 80 participants were used. For each video, 2D hand coordinate sequences were extracted using MediaPipe, followed by movement sequence computation, feature extraction, and selection. Clinicians provided severity scores as the gold standard. A LightGBM regressor was trained and evaluated using the resulting dataset of selected feature vectors and corresponding gold standard scores. Experimental results showed that our method achieved superior performance and demonstrated potential for reliable video-based evaluation.
This study addresses a task designed to predict the alignment between a natural language instruction, a pre-manipulation image, and an end-effector trajectory. Conventional methods typically perform a success prediction only after the manipulation is executed, limiting their efficiency in executing the entire task sequence. We propose a novel approach that enables the prediction of success or failure by aligning the given trajectories and images with natural language instructions. We introduce Trajectory Encoder to apply learnable weighting to the input trajectories, allowing the model to consider temporal dynamics and interactions between objects and the end effector, improving the model’s ability to predict manipulation outcomes accurately. We constructed a dataset based on the RT-1 dataset, a large-scale benchmark for open-vocabulary object manipulation tasks, to evaluate our method. The experimental results show that our method achieved a higher accuracy than the baselines.
We propose Integrated Gradients and Object Detector Activation Maps (IG-ODAM), a novel model that integrates the strengths of ODAM and Integrated Gradients to enhance the interpretability of object detection models. While ODAM effectively visualizes the decision-making processes of object detectors, it does not fully address the limitations of gradient-based attribution identified in prior work. To overcome this, IG-ODAM incorporates the path-integrated computation of Integrated Gradients into ODAM, yielding more accurate and robust visual explanations. This study represents the first application of the Integrated Gradients methodology to object detection, extending its success in image classification to the more complex domain of instance-aware detection. By leveraging both object specification and discriminative power, IG-ODAM produces high-quality, instance-specific visual explanations that satisfy sensitivity and implementation invariance axioms, thereby improving both interpretability and localization accuracy.
Denoising diffusion (score-based) generative models have recently achieved significant accomplishments in generating realistic and diverse high-dimensional data. These approaches define a forward diffusion process for transforming data into noise and a backward denoising process for sampling from noise via an approximated score function (i.e., score network). Analyzing the score function through the lens of conditional expectation, we notice that near the end of the denoising process, the score function has a blow-up (i.e., singularity), preventing the score network from approximating it properly. To address this issue, we introduce a regularity theorem for the score function and correspondingly propose our Training-Stabilized (TRAST) Diffusion Models. A key feature of our approach is noise manipulation, i.e. the addition of an unnoticeably small perturbation to the training images, particularly designed to prevent the score network from approximating the score function near its singularity. Subsequently, we incorporate a noise-filtering module at the end of the denoising process to refine the generated images. Training diffusion models can be regularized by early stopping and noise correction. Note that the designs of TRAST are theoretically supported by two theorems (i.e., singularity and regularity of the score function). Therefore, our method can stabilize diffusion models’ training by avoiding the singularity, and improve their performance in terms of fidelity and diversity.
In human-robot interaction, humans and robots should engage in natural dialogue by considering their respective perspectives on objects in a shared space. However, existing methods in 3D Dense Captioning do not support the generation of descriptions conditioned on arbitrary viewpoints. To address this issue, this paper proposes a method that incorporates viewpoint information, distinguishing between the target object and a reference object that defines its spatial relationship. This enables the method to adjust descriptions appropriately according to changes in viewpoint. The effectiveness of the proposed method is validated through both quantitative and qualitative evaluations.
Semi-supervised semantic segmentation techniques focus on Consistency Regulation, the Teacher-Student framework, Pseudo-Labeling, and self-training. Unsupervised Domain Adaptation (UDA) shares similarities with semi-supervised learning (SSL) regarding dataset characteristics, techniques, and loss functions. This research hypothesizes that UDA encompasses semi-supervision and proposes a UDA-Based SSL concept to enhance performance and broaden application possibilities. The proposed "Student Consistency Framework" strengthens the Teacher-Student model by incorporating consistency regularization, where the teacher model generates more accurate pseudo-labels for the student model. This framework also enhances the robustness of both the student and teacher models. Additionally, to address uniform feature perturbation, a new method called Feature Masking is proposed, which masks features at the patch level after encoding, compelling the model to infer masked information and improving its understanding of feature relationships.
We propose an efficient and generalizable method for skeleton-based action recognition using a novel representation called the Superposed Shape Subspace (SSS). Our approach encodes both structural and temporal dynamics of skeletal sequences by modeling multiple frames as a unified subspace in a high-dimensional space. Unlike deep learning methods that require large datasets and GPU resources, our method enables fast, CPU-based training and inference by comparing canonical angles between subspaces.To further improve efficiency, we introduce an optimal frame selection strategy that identifies the most informative frames, significantly reducing computational cost without compromising accuracy. Experiments on the First-Person Hand Action Benchmark show that our method achieves competitive performance (89% accuracy across 45 classes), processes over 100 frames per second on CPU, and supports one-shot learning with minimal data. These characteristics make it highly suitable for real-time and edge computing applications.
Age estimation models are widely used in applications such as cross-age recognition, facial age transformation, and content screening. However, the evaluation of existing models still relies on outdated baselines such as DEX and Face++, which are often inaccurate and unstable-particularly for younger age groups-thus limiting their practical applicability. In addition, current benchmarks mainly emphasize overall mean absolute error (MAE), while overlooking prediction stability and fairness across age groups. To address these limitations, we propose a lightweight Mixture of Experts (MoE) framework that dynamically combines vision-language models (VLMs), such as Janus Pro, with structure-oriented visual experts, including CNN-based (e.g., VGG16, ResNet50) and Transformer-based models (e.g., MiVOLO). Our framework improves age prediction stability through expert-specific bias correction and supports plug-andplay integration without retraining, enabling efficient fusion at minimal cost. Experiments on the AgeDB and Cross-Age Face (CAF) datasets demonstrate superior performance compared to state-of-the-art models, validating the effectiveness of our framework in age estimation and its potential for broader multi-modal applications.
Small object tracking remains a challenging task in computer vision due to factors such as low resolution, background clutter, occlusion, and limited appearance information. These challenges often lead to missed detections and identity switches, particularly when using conventional tracking-by-detection pipelines. In this paper, we propose an effective small object tracking framework that integrates enhanced detection and robust tracking strategies to address these limitations. Our approach fine-tunes the Collaborative Hybrid Assignments DETR (Co-DETR) with augmented data to improve its sensitivity to small-scale objects. To further refine detection quality, we apply Slicing Aided Hyper Inference (SAHI), which enables more precise localization through image slicing. The resulting high-quality detections are then passed to BoostTrack++, a state-of-the-art multi-object tracker. We name this pipeline Collaborative Hybrid Assignments and Tracking with slicing-aided hyper inference (CHAT). Experimental results on the SMOT4SB dataset demonstrate that our pipeline improves tracking performance on small objects, achieving more accurate detection and consistent identity preservation in complex visual scenes.
Airborne dust significantly affects numerous factors, making effective dust monitoring essential for understanding and mitigating its impact. Previous studies have demonstrated that dust density can be estimated from standard RGB images. However, pixel-wise estimation of dust concentrations is a complex task with many challenges. A significant challenge is the imbalance in the distribution of dust levels within images: dust can range from transparent to opaque, often resulting in a highly skewed distribution of dust intensities. This imbalance can lead to underestimation or overlooking of high dust concentration events. To address this issue, we propose a new balanced loss function, BinMSE, designed explicitly for pixel-wise visual regression to effectively handle imbalanced dust level distributions. We integrate our loss function into various dust density estimation algorithms and evaluate its effectiveness compared to other options on the Meteodata dust dataset.
License Plate Recognition (LPR) has been extensively studied in the fields of computer vision and pattern recognition, driven by the need for automated traffic surveillance and law enforcement. In this paper, we propose a license plate recognition using Supervised Domain Adaptation (SDA) of Scene Text Recognition framework. The proposed approach combines a CNN-based feature extraction backbone with BiLSTM sequence modeling, originally optimized for broad scene text recognition. We then adapt the pretrained model to license plate data under supervision, ensuring it can reliably capture the unique patterns and character arrangements found on vehicle plates. Furthermore, we analyze the impact of selectively freezing and finetuning specific stages, such as the transformation, feature extraction, sequence modeling, and prediction layers, to highlight the optimal adaptation scheme. The findings underscore the importance of domain-aware architectural choices and provide a practical pathway for deploying automated license plates recognition systems in resource-constrained environments, particularly where labeled data is limited.
