High-speed and accurate methods for chip-surface-defect detection remain a challenge in the semiconductor industry. Therefore, we propose a feature fusion and data generation-based cascade (FFDG-Cascade) detection approach. This method cascades a classification module with an object detection module. The classifier screens nondefective samples with high confidence, significantly mitigating the number of samples forwarded to the object detector, and substantially enhancing the efficiency due to classifiers’ higher operational speed than detectors. We enhance the model’s detection capability to detect small target defects by incorporating a shallow-to-deep attentional feature fusion (SDAFF) mechanism into the object detection module. In addition, we alleviate network overfitting issues by constructing a large dataset for advanced packaging chips. This dataset comprises 2270 real nondefective samples, 1241 real defective samples, and 7250 synthetic defective samples. For synthetic samples, we propose two defect generation algorithms, each satisfying the requirements of chip production in the early and subsequent stages. Evaluation results demonstrate that integrating synthetic data significantly enhances the detector’s performance by 7.49 mean Average Precision (mAP) on average. Upon incorporating these improvements into the FFDG-Cascade approach, the detection speed increases by 61.77%, while the false acceptance rate (FAR) and false rejection rate (FRR) are reduced by 80.67% and 59.93% on average, respectively. The dataset will be available at https://github.com/HiHiAllen/Chip-surface-defect-dataset .

Defects can be regarded as targets when using a target detection algorithm. Compared with conventional targets, chip defects have distinct characteristics. Their sizes are variable, and most defects are small in size. Defects lack texture features and can be viewed as anomalies relative to the background. Some defects exhibit elongated and strip-like characteristics, making the direct application of existing target detection algorithms less than ideal. In this paper, we incorporate these characteristics as prior knowledge in the design and improvement of the target detection network structure. We propose a deep learning detection network, YOLO-WDA (You Only Look Once - with Defect Attention), specifically tailored for chip defect data, using three targeted improvement methods. An anomaly attention mechanism highlights defect features by contrasting information with normal chips. An improved module for small target defects uses the Focus operation to retain more fine-grained information, combined with Ghost convolution to adjust channel redundancy and reduce network parameters. An Amoeba Convolution Detection Head (AMBC-Detect) can better capture continuous features such as curves. In experiments conducted on two chip datasets, YOLO-WDA achieved mAP (mean of average precision) scores of 65.5 and 43.2, outperforming the benchmark model, YOLOv8, by 2.7 and 4.8, respectively. Our model also outperforms other classical algorithms. Datasets are available at https://pan. baidu.com/s/1vU3hkPUYSrzVHDKgGgt1MA?pwd=1yja.

The main challenges in reconstruction-based anomaly detection include the breakdown of the generalization gap due to improved fitting capabilities and the overfitting problem arising from simulated defects. To overcome this, we propose a new method called PRFF-AD, which utilizes progressive reconstruction and hierarchical feature fusion. It consists of a reconstructive sub-network and a discriminative sub-network. The former achieves anomaly-free reconstruction while maintaining nominal patterns, and the latter locates defects based on pre- and post-reconstruction information. Given defective samples, we find that adopting a progressive reconstruction approach leads to higher-quality reconstructions without compromising the assumption of a generalization gap. Meanwhile, to alleviate the network’s overfitting of synthetic defects and address the issue of reconstruction errors, we fuse hierarchical features as guidance for discriminating defects. Moreover, with the help of an attention mechanism, the network achieves higher classification and localization accuracy. In addition, we construct a large dataset for packaging chips, named GTanoIC, with 1750 real non-defective samples and 470 real defective samples, and we provide their pixel-level annotations. Evaluation results demonstrate that our method outperforms other reconstruction-based methods on two challenging datasets: MVTec AD and GTanoIC.

The analysis of dynamic fluorescence diffuse optical tomography (D-FDOT) is important both for drug delivery research and for medical diagnosis and treatment. The low spatial resolution and complex kinetics, however, limit the ability of FDOT in resolving drug distributions within small animals. Principal component analysis (PCA) provides the capability of detecting and visualizing functional structures with different kinetic patterns from D-FDOT images. A particular challenge in using PCA is to reduce the level of noise in D-FDOT images. This is particularly relevant in drug study, where the time-varying fluorophore concentration (drug concentration) will result in the reconstructed images containing more noise and, therefore, affect the performance of PCA. In this paper, a new linear corrected method is proposed for modeling these time-varying fluorescence measurements before performing PCA. To evaluate the performance of the new method in resolving drug biodistribution, the metabolic processes of indocyanine green within mouse is dynamically simulated and used as the input data of PCA. Simulation results suggest that the principal component (PC) images generated using the new method improve SNR and discrimination capability, compared to the PC images generated using the uncorrected D-FDOT images.