CPD-YOLO: a chip packaging defect detection network based on attention feature orthogonal fusion and wavelet transform

Abstract Chip packaging defect detection is critical for ensuring system operational stability. Addressing the low efficiency and poor generalization of traditional manual visual inspection and early machine vision detection, and considering the challenges posed by the small size and subtle characteristics of common chip packaging defects, this paper proposes a CPD-YOLO algorithm based on YOLOv8n. Firstly, an overparameterized feature extraction module (C2f-DO) is designed to enhance the model’s feature extraction capability; secondly, an attention feature orthogonal fusion module is introduced to orthogonally integrate global and local semantic information across multiple scales, thereby strengthening the discriminative representation of small defects. Subsequently, a Haar wavelet downsampling and upsampling module is constructed to enhance the resolvability of faint defect signals in the frequency domain. Finally, we introduce InterpIoU regression loss to improve the stability and accuracy of bounding box regression for small-scale targets. On our self-built chip packaging defect dataset, CPD-YOLO achieves a mAP of 96.6%, representing a 3.9% improvement over baseline models. Concurrently, model parameters and computational complexity are reduced by 26.6% and 44.4%, respectively. Furthermore, experiments on other datasets demonstrate that CPD-YOLO outperforms alternative detection methods, validating the superiority and practicality of the proposed approach.