Application of hyperspectral imaging technology in FPCB surface defect detection

 I. Limitations of traditional visual inspection

Flexible Printed Circuit Boards (FPCB) are widely used in fields such as smartphones, flexible displays, and wearable devices due to their good bendability and heat dissipation capabilities. As circuit density continues to increase, the types of surface defects are becoming increasingly complex, with common defects including short circuits, open circuits, protrusions, white spots, black spots, and broken holes.

In traditional detection methods, template matching based on RGB images is widely used. This method locates abnormal areas by comparing a standard image with the image under test. However, these methods are sensitive to lighting conditions; when the light distribution is uneven, it is easy to produce false detections or missed detections. In addition, some defects are morphologically similar to normal circuit structures, making it difficult to distinguish them accurately relying solely on visible light images.

II. Construction of the hyperspectral imaging system

To improve the stability of detection, this study built a hyperspectral microscopic imaging system. The system consists of a hyperspectral camera, a microscope, and acquisition software. Among them, the hyperspectral camera adopts the FS-23 model from CHNSpec, which features a spectral range of 400–1000nm and a spectral resolution of 2.5nm.

The camera uses a line-scanning method for imaging, and the raw data contains 1200 bands. To facilitate processing, every four adjacent bands were merged into one in the study, finally obtaining a data structure of 300 bands. The size of a single hyperspectral image is 1920 × 960 pixels × 300 bands, covering the complete spectral information of the copper conductor and the polyimide substrate.

The advantage of hyperspectral imaging lies in its ability to obtain a continuous spectral curve for each pixel. The study found that there are significant differences in the spectral response of copper and polyimide in the 500–750nm wavelength range, which provides a reliable basis for subsequent image segmentation and material identification.

III. Spectral information-driven detection method

The detection framework proposed in this study consists of two sub-networks: FPCB-LocNet for defect localization and FPCB-ClaNet for defect classification.

In the localization stage, FPCB-LocNet utilizes multi-scale 3D convolution kernels to extract features from both spatial and spectral dimensions simultaneously. Two different sizes of convolution kernels are used in the network to focus on local spatial structures and spectral features respectively, and features of different scales are fused through a residual structure. This design allows the network to capture fine spatial textures and continuous spectral changes at the same time, achieving pixel-level segmentation of copper and polyimide. After segmentation is completed, abnormal areas are located through template matching.

In the classification stage, considering the limited number of hyperspectral samples, the network adopts a transfer learning strategy, first pre-training on the FPCB RGB image dataset and then fine-tuning on pseudo-color images. Aiming at the problem of unbalanced sample numbers for different defect categories, category-balanced sampling and weight decay strategies are introduced in the network to enable the model to focus more on defect types with fewer samples. At the same time, the SE attention mechanism is embedded to enhance the network's focus on key features.

IV. Experimental results and application value

In terms of image segmentation, FPCB-LocNet performs better than traditional segmentation methods such as entropy method, watershed algorithm, and Otsu when processing images with uneven lighting, with a segmentation accuracy reaching 97.86%. In the classification task, the comprehensive classification accuracy of FPCB-ClaNet for six common types of defects is 97.84%.

Ablation experiments verified the actual contribution of each module: data augmentation improved classification accuracy, category-balanced sampling and weight decay effectively improved the recognition effect of tail categories, and the SE attention mechanism brought a stable improvement in classification performance while adding a small number of parameters. The visualization results of Grad-CAM heatmaps show that the model's areas of concern are highly consistent with the actual defect locations.

This study combines hyperspectral imaging with deep learning to build a complete processing chain from data acquisition, image segmentation, and defect localization to defect classification. This method can stably complete the identification task of FPCB surface defects without relying on specific lighting conditions, providing a feasible technical path for the manufacturing quality management of high-density flexible circuit boards.

Product Recommendation: FigSpec FS-23 Imaging Hyperspectral Camera

• Image Resolution: 1920*1920
• Spectral Range: 400-1000nm
• Spectral Resolution (FWHM): 2.5nm
• Number of Spectral Channels: 1200