Neural network for finding mathematical formulas in videos from data science conferences

This paper presents a novel approach to the problem of finding mathematical formulas on video frames. Distinguishing equations from other text and objects on given image involves a deeper understanding of individual symbols, patterns, and their relative positions. Standard segmentation methods used for this kind of task will fail immediately because they mostly rely on noticeable differences between color and shape of different objects. By using a fully convolutional neural network and image to image transformation, we were able to achieve state of the art results in finding mathematical formulas in movie frames extracted from data science conferences videos. Current status of the work, results and further development plans are presented.