In the manufacturing of chips, the accurate and effective detection of internal bubble defects of chips is essential to maintain product reliability. In general, the inspection is performed manually by viewing X-ray images, which is time-consuming and less reliable. To solve the above problems, an improved bubble defect detection model YOLO-Xray based on the YOLOv5 algorithm for chip X-ray images is proposed. First, the chip X-ray images are preprocessed by image segmentation to construct the chip X-ray defect dataset, namely, CXray. Then, in the input stage, the K-means++ algorithm is used to re-cluster the CXray dataset to generate the anchors suitable for our dataset. In the backbone network, a micro-scale detection head is added to improve the capabilities for small defect detection. In the neck network, the bi-direction feature fusion idea of BiFPN is used to construct a new feature fusion network based on the improved backbone to fuse the semantic features of different layers. In addition, the Quality Focal Loss function is used to replace the cross-entropy loss function to solve the imbalance of positive and negative samples. The experimental results show that the mean average precision (mAP) of the YOLO-Xray algorithm on the CXray dataset reaches 93.5%, which is 5.1% higher than the original YOLOv5. Meanwhile, the YOLO-Xray algorithm achieves state-of-the-art detection accuracy and speed compared with other mainstream object detection models. This shows the proposed YOLO-Xray algorithm can provide technical support for bubble defect detection in chip X-ray images. The CXray dataset is also open and available at CXray.

Integrated circuit chips are the most common electronic components, and visual internal defect detection is essential for ensuring product quality following packaging. However, efficient detection of chip internal defects is challenging due to the complexity of the background and the faintness of defects. To overcome the above difficulties, a novel deep learning-based defect segmentation framework is proposed, which relies on an image pre-processing (IPP) scheme and a defect segmentation network (DSNet). The IPP is composed of a rotation correction algorithm and a region segmentation algorithm for removing the influence of background and obtaining chip packaging region. The DSNet is proposed to precisely and efficiently segment the internal defects. To address the scarcity of data and avoid overfitting, we proposed the lightweight convolution block by using depth-wise separable convolution to reduce the number of parameters. Besides, the attention gate module is incorporated into the skip connection to handle the shape varieties of the defects. Moreover, the Focal loss function is designed to guide the network to pay attention to small defects that are difficult to distinguish. The robustness and adaptability of the proposed method are evaluated on three typical types of chip X-ray datasets from real-world inspection lines. Experimental results show that the proposed framework achieves a satisfactory trade-off between detection accuracy and speed with an F1 score of 72.69% and an FPS of 17.5 on average, resulting in superior segmentation performance even with sparse or insufficient data.

Remote-sensing object detection commonly requires repeated training on high-resolution aerial and satellite imagery, where targets may be small, densely distributed, and surrounded by extensive background. For remote-sensing detection tasks that require shorter model-training cycles, reducing training time without sacrificing detection accuracy is important. The Anti-Forgetting Sampling Strategy (AFSS) reduces training time by avoiding repeated processing of learned images, but its fixed warm-up may start sampling before the detector is mature and thereby reduce accuracy. We propose Maturity-Gated AFSS (MG-AFSS), a detector-maturity activation controller for AFSS. The method accumulates validation mean average precision at an intersection-over-union threshold of 0.50 (mAP50), fits a cumulative saturating curve online, and enables AFSS only when the estimated maturity indicates trustworthy image states. After activation, it reuses the original AFSS sampling rule. We evaluate MG-AFSS with matched lightweight detectors on five public remote-sensing datasets: NWPU VHR for the main three-seed study and additional settings; UCAS-AOD for horizontal-box detection; HRSC2016 and ShipRSImageNet for oriented bounding-box (OBB) boundary cases; and DOTAv1 for a larger-scale OBB block. In the three-seed, 80-epoch, from-scratch NWPU VHR experiment, MG-AFSS improves over AFSS by +0.0264 mAP50 and +0.0319 mAP50–95, where mAP50–95 denotes mean average precision averaged over intersection-over-union thresholds from 0.50 to 0.95. It keeps mAP50–95 essentially equal to standard full-dataset training (+0.0003), with a training-time ratio of 0.867 relative to standard training. In a pretrained NWPU setting, the training-time ratio is 0.557 relative to matched standard training, with a higher checkpoint mAP50. In the DOTAv1 block, MG-AFSS has a training-time ratio of 0.759, lower than AFSS at 0.777; its checkpoint-best differences from standard training are −0.0095 mAP50 and −0.0149 mAP50–95, both smaller in magnitude than those of AFSS. Overall, MG-AFSS yields a conditional accuracy–training-time trade-off: it reduces training time when the maturity condition is satisfied and, as observed in the two smaller OBB boundary runs, retains full-dataset training otherwise.