Long-form data, such as video or text, often encompasses a multitude of features, each exerting varying degrees of influence on the final categorization outcome within any given segment of the data. In this paper, we propose a Segmented Augmentation Method tailored for small-scale long-form data. This method leverages k-means clustering to categorize each segment of the data based on its distinct features, employing a two-level classification approach. Subsequently, data frames wherein all features align with a single class are selected to construct a training set. This curated training set facilitates the discernment of human preferences. The DNNs with an overlay mechanism is then trained utilizing both segmented and unsegmented training sets specific to a given dataset, resulting in a notable enhancement in network performance.