The process of image binarization is widely utilized in image processing, however, in embedded systems, yet traditional algorithms exhibit output instability. To address this issue, a method for optimizing the image binarization threshold based on Gaussian weighted averaging and machine learning is proposed in this paper. In contrast to traditional algorithms, the classic Otsu method is employed to determine the binarization threshold, and Gaussian weighted averaging and machine learning techniques are introduced to smooth, denoise, and limit the threshold, thereby enhancing the stability of the binarization algorithm in embedded systems. To validate the effectiveness of the algorithm, threshold data collection was conducted on an embedded system using the TC264 chip, and a detailed comparative analysis of the results from various binarization algorithms was performed. The experimental results demonstrate a significant improvement in both stability and accuracy of the algorithm.