Vision–language models (VLMs) have achieved remarkable success in natural image domains, yet their potential remains underexplored in agriculture due to the lack of high-quality, joint image–text datasets. To address this limitation, we introduce VL-PAW (Vision–Language dataset for Pear, Apple, and Weed), a dataset comprising 3.9 K image–caption pairs for two key agricultural tasks: weed species classification and fruit inspection. We fine-tune the CLIP model on VL-PAW and gain several insights. First, the model demonstrates impressive zero-shot performance, achieving 98.21% accuracy in classifying coarse labels. Second, for fine-grained categories, the vision–language model outperforms vision-only models in both few-shot settings and entire dataset training (1-shot: 56.79%; 2-shot: 72.82%; 3-shot: 74.49%; 10-shot: 83.85%). Third, using intuitive captions enhances fine-grained fruit inspection performance compared to using class names alone. These findings demonstrate the applicability of VLMs in future agricultural querying systems.

Recent approaches utilizing self-supervised learning with masked image modeling (MIM) have demonstrated great performance. However, applying MIM naively to small datasets results in poor generalization to downstream tasks. We hypothesize that capturing detailed anatomical structures can compensate for the limitations posed by the dataset shortage; thus, we introduce Spectral Distance Scaling Loss (SDSL), designed to improve generalization in medical imaging tasks. Unlike traditional pixel-based methods, SDSL incorporates frequency-domain information to enhance encoder representations, ensuring balanced learning of both low and high-frequency details. Furthermore, wavelet multi resolution decomposition was utilized to enable the pretrained model to reconstruct frequency information across multiple stages. The comprehensive experiments demonstrate that SDSL pretraining yields sharper reconstruction results and more accurate segmentation outcomes than existing methods. The proposed approach achieved the highest average Dice scores of 84.17% on the Beyond the Cranial Vault dataset, 98.20% on the Medical Segmentation Decathlon Spleen dataset, and 90.38% on the Multimodality Whole Heart Segmentation dataset. The findings highlight the potential of SDSL in advancing medical imaging techniques by effectively handling spectral variations and improving model generalization.

This work discusses the challenges of multi-label image classification and presents a novel Efficient Shuffle Net (EffShuffNet) based on a convolutional neural network (CNN) architecture to address these challenges. Multi-label classification is difficult as the complexity of prediction increases with the number of labels and classes, and current multi-model approaches require optimized deep learning models which increase computational costs. The EffShuff block divides the input feature map into two parts and processes them differently, with one half undergoing a lightweight convolution and the other half undergoing average pooling. The EffShuff transition component shuffles the feature maps after lightweight convolution, resulting in a 57.9% reduction in computational cost compared to ShuffleNetv2. Furthermore, we propose EffShuff-Dense architecture, which incorporates Dense connection to further emphasize low-level features. In experiments, the EffShuffNet achieved 96.975% accuracy in age and gender classification, which is 5.83% higher than the state-of-the-art, while EffShuffDenseNet was even better with 97.63% accuracy. Additionally, the proposed models were found to have better classification performance with smaller model sizes in fine-grained image classification experiments.

본 연구에서는 임베디드 시스템 기반 사과 품질 예측을 위한 최적의 딥러닝 모델 탐색에 관한 연구를 수행하였다. 탐색을 위해 경량 모델 6종을 학습하고 정확도와 추론 속도에 대한 가중평가 하였다. 실제 수집한 사과 품질 이미지 데이터셋을 활용하여 전이학습을 진행하였으며, 그 결과 RegNet-y 모델이 가장 우수한 성능을 보였다. 이를 임베디드 시스템에의 배포하기 위해 최적화 실험을 거쳐, 최종적으로 테스트 데이터셋 2,000장에 대해서 정확도 95.36%와 이미지 당 0.0566초의 추론속도를 달성하였다. 본 연구의 결과는 임베디드 시스템에서 효율적인 딥러닝 모델 활용에 대한 경험적 데이터를 제공하여 실시간 애플리케이션 가능성을 확장한다.

Automated radiology report generation from chest X-rays presents a critical challenge in medical imaging. Traditional image-captioning models struggle with clinical specificity and rare pathologies. Recently, contrastive vision language learning has emerged as a robust alternative that learns joint visual–textual representations. However, applying contrastive learning (CL) to radiology remains challenging due to severe data scarcity. Prior work has employed input space augmentation, but these approaches incur computational overhead and risk distorting diagnostic features. This work presents CL with feature space interpolation for retrieval (CLFIR), a novel CL framework operating on learned embeddings. The method generates interpolated pairs in the feature embedding space by mixing original and shuffled embeddings in batches using a mixing coefficient $\lambda \sim U(0.85,0.99)$. This approach increases batch diversity via synthetic samples, addressing the limitations of CL on medical data while preserving diagnostic integrity. Extensive experiments demonstrate state-of-the-art performance across critical clinical validation tasks. For report generation, CLFIR achieves BLEU-1/ROUGE/METEOR scores of 0.51/0.40/0.26 (Indiana university [IU] X-ray) and 0.45/0.34/0.22 (MIMIC-CXR). Moreover, CLFIR excels at image-to-text retrieval with R@1 scores of 4.14% (IU X-ray) and 24.3% (MIMIC-CXR) and achieves 0.65 accuracy in zero-shot classification on the CheXpert5×200 dataset, surpassing the established vision-language models.

Foundation models, exemplified by the Segment Anything Model (SAM), have revolutionized object segmentation with their impressive zero-shot capabilities. The recent SAM2 extended these abilities to the video domain, utilizing an object pointer and memory attention to maintain temporal segment consistency. However, a critical limitation of SAM2 is its vulnerability to error accumulation, where an initial incorrect mask can propagate through subsequent frames, leading to tracking failure. To address this, we propose a novel method that actively monitors the temporal segment consistency of masks by evaluating the distance of object pointers across frames. When a potential error is detected via a sharp increase in distance, our method triggers a particle filter based re-inference module. This framework models object’s motion to predict a corrected bounding box, effectively guiding the model to recover the valid mask and preventing error propagation. Extensive zero-shot evaluations on DAVIS, LVOS v2, YouTube-VOS and qualitative results show that the proposed, parameter-free procedure consistently improves temporal coherence, raising mean IoU by 0.1 on DAVIS, by 0.13 on the LVOS v2 train split and 0.05 on the LVOS v2 validation split, and by 0.02 on YouTube-VOS, thereby offering a simple and effective route to more robust video object segmentation with SAM2.