Popular music relies on soundtrack recognition research, yet there's an issue with erroneous recognition and placement. Unfortunately, the issue of popular music soundtrack identification remains unsolved by the conventional neural network approach, and the results are far from satisfactory. Consequently, this study reviews the literature on popular music score identification and suggests future deep learning-based research in the field. In order to minimize interference elements in soundtrack recognition research, we first apply optimization theory and gradient descent to find the influencing variables. Then, we split the indicators according to the needs of soundtrack recognition research. The next step is to establish a deep learning framework for soundtrack recognition study using optimization theory and gradient descent. After that, the outcomes of this research will be thoroughly examined. In terms of accuracy and duration of influencing elements in soundtrack recognition study, the MATLAB simulation results demonstrate under specific evaluation criteria.