The deep classifiers used for chest X-ray diagnosis can be sensitive to the visual frame around the evidence, producing correct labels for the wrong reasons by relying on the background, borders, text marks, or acquisition artifacts. This framing sensitivity reduces their trustworthiness, and they can fail under a distribution shift. Inspired by biological vision, especially figure–ground segregation and selective attention, we propose a bio-inspired adversarial attention alignment training process that encourages evidence-centered decisions without changing the classifier structure. A classifier is first trained with image-level labels. Class activation mapping (CAM) is then used to produce a differentiable heatmap that indicates where the model attends. We treat this heatmap as a generated localization map and train a discriminator to distinguish the generated heatmaps from ground truth masks. The classifier is updated using a joint objective that preserves the classification performance while pushing its CAM heatmap toward a mask-like structure, reducing the reliance on background cues. We also introduce evaluation measures for the test phase, including augmentation inconsistency (prediction flip rate under angle-based augmentations) and framing sensitivity (CAM energy outside the mask). The experiments show improved lung-focused attention and robustness, while requiring masks only during training and no additional inputs at inference.

The accuracy of camera-based plant disease classification is crucial for smart farming applications, particularly in automated management systems using agricultural robots. To address the performance degradation caused by motion blur, we propose a knowledge distillation (KD)-based image deblurring network (KDID-Net). Our framework targets classification performance by integrating generative adversarial network (GAN)-based deblurring with fast Fourier transform (FFT)-based complex-domain KD. Specifically, the teacher model in KDID-Net employs a classification-guided (CG) block that generates logits analogous to those produced by a classifier, which are then used to compute a classification-guided enhancement loss (CGEL), thereby reinforcing features relevant to disease classification. Furthermore, the teacher network’s feature maps are subjected to FFT, and frequency-domain knowledge is distilled into a lightweight student network, enabling the recovery of fine lesion details. Consequently, the lightweight network achieves high classification accuracy while remaining suitable for agricultural robots with limited computational resources. Moreover, the disease classification results produced by the proposed method are automatically linked to the input of the open-source large language model Meta artificial intelligence (LLaMA), enabling a system that recommends optimal pesticide information for each disease class. We conducted experiments with two public datasets: the PlantVillage and FieldPlant datasets. Under the same blur conditions, KDID-Net achieved classification accuracies of 0.9432 and 0.7571, respectively, outperforming state-of-the-art (SOTA) image restoration methods and KD methods in terms of classification accuracy. These results demonstrate the effectiveness of the proposed method for plant disease classification under both simple and complex background conditions. Furthermore, the proposed model achieved superior performance compared with SOTA methods in terms of graphic processing unit (GPU) memory requirements, floating point operations (FLOPs), and inference times on computationally constrained embedded systems, demonstrating its suitability for agricultural robot applications.

Although deep learning methods have shown high performance in chest X-ray classification, high accuracy alone does not guarantee reliable reasoning. A model may still exhibit pathological behavior, such as unstable evidence usage under harmless input changes, inconsistent reasoning across augmented views, excessive dependence on surrounding frame information, and appearance bias dominance, where prediction relies too heavily on intensity while neglecting texture and shape. In this paper, we propose a bio-inspired pathology-aware, factor-aware framework for explainable and reliable chest X-ray classification, inspired by biological vision principles such as figure–ground separation, selective attention, and balanced use of complementary visual cues. During training, the method regularizes appearance bias dominance through evidence-guided counterfactual perturbations that mimic cue-suppression analysis in biological perception, thereby revealing and penalizing excessive factor dependence. During testing, it evaluates model behavior using four criteria: reasoning stability, augmentation inconsistency, appearance bias dominance, and framing sensitivity. This combination enables the framework to go beyond conventional inference-time explanation by both correcting pathological behavior during training and exposing it during evaluation. From a biomimetic perspective, the framework encourages the model to separate relevant foreground anatomy from surrounding background and to avoid over-reliance on a single dominant cue. The proposed approach improves interpretability and reliability without modifying the backbone architecture or increasing model size or inference-time cost. The proposed training process improved the F1-score of DenseNet-121 from 0.899 to 0.931, while also producing more stable and balanced reasoning.

Background/Objectives: Automated visual inspection of pharmaceutical tablets is increasingly important for detecting manufacturing defects and maintaining consistent product quality. Progress in artificial intelligence-based defect classification is, however, constrained by the scarcity of publicly accessible image datasets covering defective tablets under diverse visual conditions. This study introduces the PharmaVision (PV) dataset, a public visual dataset and benchmark for classifying normal and defective compressed tablets. Methods: The PV dataset contains 640 smartphone-acquired images spanning varied tablet products, backgrounds, and lighting, with fixed tablet-level splits (190/190 training, 30/30 validation, and 100/100 test images for the normal and defective classes, respectively). Twelve representative architectures spanning classical convolutional networks, mobile networks, and vision transformers were benchmarked under an identical protocol using accuracy, precision, recall, and harmonic mean of precision and recall (F1-score), with 95% confidence intervals and paired statistical tests. Cross-dataset generalization was assessed on the independent pillQC dataset without fine-tuning, and efficiency was profiled using parameters, Giga floating-point operations (GFLOPs), latency, and throughput. Results: EfficientNetV2-S achieved the highest F1-score on the PV test set (88.56%), although paired statistical tests did not distinguish it from Xception (87.08%) or Swin-B (85.71%). On pillQC, GoogLeNet and EfficientNetV2-S showed the strongest cross-dataset performance, whereas both vision transformers declined relative to their in-domain results. AlexNet gave the highest theoretical throughput (602.41 frames per second (FPS), median latency 1.66 ms), whereas ShuffleNetV2 and MobileNetV3-small offered the smallest parameter counts. Conclusions: In-domain accuracy, cross-dataset robustness, and measured latency diverged, with no single architecture best on all three axes. The PV dataset, code, and evaluation protocol are publicly available on GitHub to support reproducible AI-based tablet quality control.