The primary aim of this study is to address the challenges of imaging difficulties and low contrast in detecting extremely fine transparent appearance defects in polarizing films. Polarizing films are common optical elements widely used in applications such as liquid crystal display panels, which include two layers of polarizers with orthogonal polarization directions. A polarizing film typically consists of six micrometer-thick transparent polymer films and aesthetic defects can occur in any of these layers. Defects such as bumps, foreign objects, bubbles, and scratches can directly reduce the quality grade of display panels, even leading to the scrapping of the entire panel. Therefore, research on detection technologies for aesthetic defects in polarizers holds significant practical importance. While most current studies focus on algorithm improvements, less attention has been given to imaging methods. In this paper, we propose a new defect detection method based on polarimetry basis parameters (PBPs) imaging. The proposed method enhances the accuracy of defect detection and classification by acquiring the Mueller matrix of defective samples, performing matrix decomposition and transformation to derive a series of PBPs with clear physical meanings, and utilizing the PBPs image with the best defect imaging contrast. More importantly, the PBPs of defects can provide rich and comprehensive information, which is expected to help polarizer manufacturers analyze the causes of defects and adjust production line processes promptly.

The Mueller matrix of the sample is captured using the fundamental measurement method for Mueller matrices. A set of PBPs, which describe the polarization characteristics of the sample, is derived through Lu-Chipman polar decomposition and several matrix transformations. After analyzing the distribution curves of the defect PBP values, comparing the PBP values between defective and non-defective areas, and considering the computational speed, we select the linear diattenuation characteristic as the outcome of the polarimetry basis parameter imaging (PBPI) method. This characteristic is utilized to differentiate common aesthetic defects from non-defective regions. We then compare the PBPI method with the traditional polarizer-sample-analyzer (PSA) imaging method, demonstrating that the PBPI method can significantly enhance the imaging contrast of minor defects. To address the scarcity of defect samples, we employ a stable diffusion model with the low-rank adaptation (LoRA) method for parameter-efficient fine-tuning to augment the data. Finally, a lightweight network, MobileNetV3_small, is used for the detection of polarizing films. The model’s performance is evaluated using precision, recall, and F1-score metrics.

The experimental results indicate that the PBPs of defects generally conform to a normal distribution (Fig. 2), allowing the use of average values from multiple samples of the same category to characterize this category. Among the 11 PBPs that are independent of rotation angle, the diattenuation and linear diattenuation characteristics show the greatest distinction between defective and non-defective areas (Fig. 4) and perform well in terms of NIQE and BRISQUE metrics (Fig. 5). However, the calculation formula for linear diattenuation involves one less parameter compared to the diattenuation characteristic (Equations 3 and 4), making it more suitable for defect detection in polarizers. By combining four special incident lights, the calculation formula for linear diattenuation can be further derived (Equation 8). Using this method to obtain the linear diattenuation characteristics of the sample reduces the average measurement time from 134.25 to 32.22 s, achieving a more than fourfold increase in speed, making it significantly more efficient than the original method. In addition, compared to the traditional PSA imaging method, the PBPI method significantly enhances contrast, especially for transparent defects such as bubbles and indentations, where the contrast improves by 2?7 times (Table 2). In the detection results using lightweight networks, the PBPI method’s detection rates for scratches, bubbles, indentations, and non-defective samples are all higher than those of the PSA method (Table 3). Furthermore, in terms of evaluation metrics such as precision, recall, and F1-score, the PBPI method outperforms the PSA method (Table 4), demonstrating its effectiveness.

In this paper, we introduce a novel method, PBPI, for detecting aesthetic defects in polarizers using PBPs. The method capitalizes on the linear diattenuation characteristic of defects to enhance their visibility and facilitate detection, benefiting quality control of polarizing films. Compared to the PSA method, PBPI method significantly improves the ability to distinguish between defective and non-defective areas. We have refined the calculation formula for linear diattenuation, reducing computational load and increasing measurement speed by four times. To address the challenge of limited defect samples, we employ a stable diffusion model and LoRA fine-tuning method for data augmentation, effectively expanding the dataset and enabling robust training even when defect samples are scarce. Finally, we utilize a lightweight network, MobileNetV3_small, for detection, achieving an average detection rate of 99.3%. In summary, the PBPI method is an effective imaging technique, particularly for detecting subtle defects in polarizing films. Future work will focus on further optimizing the method for real-time applications and exploring its potential in other areas of industrial inspection.