張檢測：基於捲積變分自動編碼器快速重建的血糖試片瑕疵檢測

Abstract

血糖酶试片是用以检测血糖水平的医学试纸，它在目前的工业生产中虽然以及完成自动化，但是其瑕疵品检测的部分还是需要大量的人力进行目检，人力检测效率和检测率都不稳定，最终会影响工厂的产量和良品率。 所以开发一款自动化的检测设备刻不容缓。但是工业瑕疵检测都面临着瑕疵样本难以获得的问题，而血糖酶试纸瑕疵检测更是面临检测速度要求快，检测精度高，产品类型多样和瑕疵样式不固定等特点，基于这些，我们开发了一款基于VAE[3]的捲積VAE和基於SSM的損失函數算法的ZhangInspect血糖酶試片瑕疵檢測算法，通過變分自動編碼器的圖像重建和去噪功能，巧妙地通過補全瑕疵樣本以找到瑕疵位置，并且通過捲積的方法提取空間結構信息，使得重建結果更加精確。通過以上方法，相較於傳統的重建瑕疵檢測，模型訓練時間和重建時間都大大下降，準確率也非常高，增加了算法的實用性和適用性。

Glucose enzyme test strips are medical test strips used to detect blood sugar levels. Although they have been automated in current industrial production, the anomaly detection part still requires much manpower for visual inspection. The efficiency and detection rate of manual inspection are unstable, which will eventually affect the factory's output and yield rate. Therefore, it is urgent to develop an automated inspection device. However, industrial anomaly detection faces the problem of difficulty in obtaining anomaly samples, and glucose enzyme test strip anomaly detection faces the characteristics of fast detection speed, high detection accuracy, diverse product types, and non-fixed anomaly patterns. Based on these, we develop ZhangInspect glucose enzyme test strip anomaly detection algorithm based on the convolution Variational Auto-Encoder (VAE) [3] and the loss function algorithm based on Sum of Square Error (SSE). Through the image reconstruction and denoising function of the variational autoencoder, the anomaly location is found by ingeniously completing the anomaly sample, and the spatial structure information is extracted through the convolution method, making the reconstruction result more accurate. Through the above methods, compared with traditional reconstruction anomaly detection, ZhangInspect training time and reconstruction time are greatly reduced, the accuracy is also very high, and the practicality and applicability of the algorithm are increased.