Automated crack recognition has achieved remarkable progress in the past decades as a critical task in structure health monitoring, to ensure safety and durability in many industrial scenarios. However, imbalanced crack recognition remains challenging due to the scarcity of crack samples and the consequential limited diversity. To resolve this, Artificial Intelligence Generated Content (AIGC) has been gradually adopted to generate synthetic data and reduce reliance on large amounts of labeled crack samples. This paper assumes that a crack sample in the feature space can be regarded as a combination of crack and background semantics. Then, the decompose-compose feature augmentation framework (DeCo) is proposed to perform crack data synthesis in the feature space by randomly composing crack and background semantic-relevant features. Specifically, the contrastive learning-based decomposing loss is proposed to enforce two encoders to separately learn crack and background semantics from crack samples with the theoretical guarantee. After that, an effective cross-instance feature union strategy is proposed to synthesize diverse crack samples by composing the crack-relevant features from a crack sample and background-relevant features across other training samples. To address the limited availability of related benchmarks, we collect INPP2022 and IRC2022 datasets from real-world applications in nuclear power plants and road pavement. Experimental results show that DeCo performs favorably against state-of-the-art competitors in imbalanced crack recognition tasks.

Imitation learning (IL) shows its superiority in faster strategy optimization by leveraging expert demonstrations. However, as one of the powerful IL methods, behavior cloning (BC) suffers from covariate shift problems, where the agent’s policy drifts away from the expert’s, leading to compounding errors and decreased generalization performance. To this end, we propose Search-based Inverse Dynamics Imitation Learning, namely SIDIL, to enhance the robustness of imitation learning by augmenting expert demonstrations via trajectory perturbation and stitching. Specifically, SIDIL first employs a nearest-neighbor search method to find the closest points between expert and perturbed data, generating new actions near the expert data via a stitching strategy. By exploiting this, the agent can recover from deviations and complete tasks under a wider range of conditions. Experimental results on various robotic tasks show that SIDIL outperforms baseline algorithms with higher success rates across multiple tasks. Meanwhile, SIDIL is allowed to expand the attractive region around expert demonstrations by stitching states from expert demonstrations, enjoying a higher task completion success rate even for states outside the expert distribution. These results highlight our potential to enhance humanoid locomotion and dexterous robotic hand manipulation, making it particularly suitable for industrial manufacturing automation where adaptability, robustness, and safety are essential.

Offline imitation learning (IL) enables embodied agents of AI-driven automation control systems to acquire policies from demonstrations without interacting with the environment. However, collecting extensive expert demonstrations that rely on a manually designed reward function is labor-intensive, and even impractical for dynamic real-world settings. To mitigate the reliance on large amounts of reward-annotated expert data, previous works have made great efforts by leveraging both expert and suboptimal demonstrations. In this paper, we reveal that existing methods still suffer from suboptimal results for two reasons: (i) reward signals in suboptimal demonstrations are not reliable, due to these rewards derived from inexperienced annotators, thus can not reflect the actual rewards, and (ii) those suboptimal demonstrations, which are significantly distant from expert behaviors, result in low reconstructed rewards. Moreover, motivated by the recent diffusion-based foundation model, we propose Expert-Like Transition Imitation Learning (ELTIL), a novel generative trajectory augmentation method that effectively leverages diffusion models. ELTIL re-designs the reward function based on expert proximity, and then proposes amplified return conditioning for achieving high-quality trajectory generation. More specifically, instead of relying on environment-provided rewards or manually designed rewards, ELTIL first serves as a good reward labeler for transitions by measuring their distance to expert states. These re-designed rewards then serve as guidance in the diffusion process, with amplified returns contributing as conditioning values to bias generation toward expert-consistent behaviors. Extensive experiments on diverse D4RL benchmarks demonstrate that ELTIL consistently improves the performance and stability of reward learning and offline IL methods, maintaining reliability under varying noise levels, reward sparsity, and dataset distributions. These characteristics make ELTIL well-suited for deployment in robust and reliable automated control systems such as robotic manipulation and autonomous driving.

Crack segmentation receives extensive attention in structure health monitoring for many industrial scenarios, e.g., bridges, highways, and nuclear power plants. The current deep learning-based crack segmentation models enjoy the ability to extract discriminative crack features by training with an extensive labeled crack dataset. However, collecting extensive crack samples with accurate annotations from experts for a new scenario is labor-intensive, thereby limiting the effectiveness of these deep models in practical applications. To address this problem, the existing Fourier-based augmentation adopts a vanilla amplitude fusion process, i.e., the portion of amplitude components is fixed or randomly selected, failing to guarantee augmented samples’ semantics consistency, and diversity concerning the original sample. To fill this, this article proposes a self-adaptive Fourier augmentation framework that efficiently synthesizes diverse crack samples for training crack segmentation models. Our proposed framework advances Fourier transformation in an adversarial learning manner, alternating between self-adaptive Fourier-based data augmentation and teacher–student learning. The former aims to guarantee the diversity and semantics consistency of Fourier-based augmented samples, while the latter progressively updates the student network by observing these augmented samples for extracting discriminative features via a knowledge distillation mechanism. It is worth noting that the proposed method is only applied in the training stage without extra computation and memory during inference. Extensive experiments demonstrate the superiority of our method over the existing methods.

The detection of optic disc and macula is an essential step for ROP (Retinopathy of prematurity) zone segmentation and disease diagnosis. This paper aims to enhance deep learning-based object detection with domain-specific morphological rules. Based on the fundus morphology, we define five morphological rules, i.e., number restriction (maximum number of optic disc and macula is one), size restriction (e.g., optic disc width: 1.05 +/- 0.13 mm), distance restriction (distance between the optic disc and macula/fovea: 4.4 +/- 0.4 mm), angle/slope restriction (optic disc and macula should roughly be positioned in the same horizontal line), position restriction (In OD, the macula is on the left side of the optic disc; vice versa for OS). A case study on 2953 infant fundus images (with 2935 optic disc instances and 2892 macula instances) proves the effectiveness of the proposed method. Without the morphological rules, naïve object detection accuracies of optic disc and macula are 0.955 and 0.719, respectively. With the proposed method, false-positive ROIs (region of interest) are further ruled out, and the accuracy of the macula is raised to 0.811. The IoU (intersection over union) and RCE (relative center error) metrics are also improved .

Retinopathy of Prematurity (ROP) is a retina disorder that affects premature infants with lower weights. If the patient cannot get the treatment in time when the illness reaches the last stage, irreversible vision loss will be caused. Nevertheless, there has been relatively little consideration given to the segmentation of the ridge, the key clinical characteristic of the illness. Additionally, existing research has not adequately addressed several segmentation issues, such as fragmentary topology, class imbalance, and false positives. This paper proposes a Densely Dilated U-Net (DD-UNet) improved from U-Net to tackle these challenges. Furthermore, the post-processing techniques based on the spatial relationship between vessels and ridges, along with the relative pixel counts of ridges and false positive results is integrated to mitigate false positive results in the predicted ridge. To enhance the precision of thin vessel, a sliding window sampling method is introduced for refined training. Compared with the state-of-the-art models in medical image segmentation, DD-UNet performs well in curvilinear structure segmentation of fundus image. For instance, our DD-UNet outperforms the Attention U-Net by 6.26% in terms of sensitivity and exhibits a 1.85% higher dice score in ridge segmentation.

Auscultation is an important tool for diagnosing respiratory-related diseases. Unfortunately, the quality of auscultation is limited by the professional level of the doctor and the environment of the auscultation. Some studies have focused on automated auscultation techniques. However, existing approaches suffer from two challenges: 1) the models cannot learn from data distributed among multiple hospitals and 2) the predictions of the models are difficult to interpret for physicians. To address this issue, this article proposes a novel explainable respiratory sound analysis framework with fuzzy decision tree regularization. This framework develops an ensemble knowledge distillation technique to learn distributed data and achieves good performance in terms of model efficiency and accuracy. Fuzzy decision trees are used to explain the predictions of the model and produce decision rules that can be well accepted by physicians. The effectiveness of this framework is thoroughly validated on the Respiratory Sound database and compared with other existing approaches.