In semiconductor chip manufacturing, defect detection is a crucial step for ensuring product quality and process stability. However, obtaining defect images in actual production is highly challenging, as most images are defect-free. Such data imbalance may result in missed or incorrect detections in downstream tasks, thereby adversely affecting the semiconductor chip's fabrication process and its final quality. To address this issue, we propose the "Spatial Anomaly Embedding Diffusion (SAE-Diffu)" model. This model incorporates domain knowledge of semiconductor chip fabrication and integrates a spatial anomaly embedding module. It generates defect images that exhibit both domain realism and diversity, while also enhancing the focus on small-target defects and enabling controllability of the defect regions. Furthermore, to improve the model's efficiency, we introduce a condition-guided single-step denoising mechanism. This mechanism reduces the number of iterations while ensuring the quality of the generated defect images is maintained. The experimental results clearly demonstrate that SAE-Diffu generates defect images which are superior in terms of realism and diversity when compared to existing methods. This serves as strong validation of the model's effectiveness and applicability within the semiconductor domain.

To solve the problem that the depth-of-field limitation of optical imaging systems leads to the inability of a single image to capture the clarity of the entire scene, we propose FADFNet,a Focus Attention and Dual-Feature Adaptation Multi-Focus Image Fusion Network. Traditional fusion methods rely on handcrafted features such as gradients or contrasts, which usually make it difficult to distinguish between focused and defocus areas in complex scenes, while deep learning methods often produce artifacts or lose details in transition areas.FADFNet overcomes these challenges by integrating three key components: the Focus Attention (FA) block, which captures pixel-level global dependencies and dynamically adjusts feature importance through an adaptive attention mechanism; the Dual-Feature Adaptive (DA) block, which employs a bidirectional structure to globally weight and embed focused pixels, thereby enhancing focused region features; and the Feature Pyramid Network (FPN), which modulates near- and far-focus feature maps to produce a high-resolution fused images. The experimental results show that FADFNet is superior to existing methods in terms of fusion quality.

Fabric defect detection plays a pivotal role in achieving intelligent quality control within textile manufacturing. However, intricate background textures, the high visual similarity between defects and the background, and the low proportion of small defects in high-resolution images impede detection accuracy. To address these challenges, we introduce Fabric-DETR, an efficient model for multi-scale fabric defect detection in complex environments. This network integrates three key modules: the Bottle Neck Conv2X Block for enhanced backbone feature extraction, Dynamic Attention-based Intra-scale Feature Interaction to improve attention to small targets, and the Zoom Diffuse Pyramid Network for efficient multi-scale feature fusion. Experiments on a six-class fabric defect dataset demonstrate that Fabric-DETR outperforms existing state-of-the-art methods, achieving a 94.8% mAP, representing improvements of 2.8%, 2.8%, 1.9%, 6.1%, 3.1%, and 2.5% over RT-DETR, YOLOv5-m, YOLOv8-m, YOLOv10-m, YOLOv11-m, and YOLOv12-m, respectively.

The Segment Anything Model (SAM) lacks adaptability to specific industrial domains, hindering its ability to leverage process-specific knowledge for adaptive optimization. Furthermore, its performance on segmenting low-contrast targets is suboptimal, causing blurred or imprecise boundary delineation of subtle defects. Finally, the reliance on manual prompt inputs limits its suitability for the fully automated, high-efficiency inspection requirements of industrial production. To address these limitations, this study introduces the Industrial Defect-SAM (ID-SAM) for industrial defect segmentation. Initially, we incorporate expert knowledge and industrial process data to construct a structured prior knowledge base of defect patterns. This knowledge is then transformed into vectors via a Contrastive Language-Image Pretraining (CLIP) mechanism, enabling the model to utilize domain-specific knowledge for adaptive optimization. Subsequently, we employ the global feature fusion, integrating the prior knowledge vectors generated by CLIP with the raw features from the image encoder. This joint modeling approach, leveraging cross-modal alignment, enhances the segmentation accuracy of low-contrast targets. Finally, a CLIP-based semantic mechanism is designed that utilizes cross-modal semantic understanding to enable the model to automatically perceive target regions, eliminating the need for manual prompts, thereby improving inspection efficiency. Experimental results demonstrate that the ID-SAM effectively enhances the segmentation accuracy of subtle defects and improves generalization capabilities across different process environments in semiconductor chip defect detection tasks.