The collection and archiving of electrocardiogram (ECG) images in the real world are critical issues in both the field of imaging science and cardiology. Statistical analysis of large-scale ECG images can yield meaningful insights; however, extracting useful information from real ECG images poses a significant challenge. In this paper, we, team "Toy-Story", present optimizations in both ECG signal extraction and classification. Firstly, we propose a denoising scheme for ECG images based on a low-rank approximation of the feature matrix. By employing random projection and autocorrelation-based low-rank approximation, we achieve substantial improvements in denoising speed while significantly reducing image storage space, without compromising denoising effectiveness. Subsequently, we develop a framework for signal extraction and compensation of missing information, aiming to maximize the extraction of signal information from the original images. Finally, we construct a multi-view deep neural network model that learns and integrates features from multiple leads, achieving highly desirable classification performance.