With the development of educational technology, there are problems such as feedback lag, subjective evaluation, and training fragmentation in traditional piano art instruction. This paper explores the influence of realtime feedback generation algorithm on the improvement of playing skills in piano art instruction. The study designs and implements a deep learning-based piano hand fingering recognition system, which uses the YOLOv3 target detection algorithm to identify playing errors, combines with the HRNet network for gesture estimation, and compares and scores the player’s playing with the standard data through the DTW algorithm. Evaluation on the FHAD dataset shows that the average joint error of this system is 16.18 mm, which is better than the comparison methods NTIS of 16.88 mm, Crazyhand of 22.03 mm and BT of 27.24 mm.When the error threshold is above 26 mm, the estimation effect of this system exceeds that of all comparison methods, which indicates that it has an excellent performance in the piano hand fingering feature extraction and fusion with excellent performance. Experiments on two piano art classes (100 students in total) in the College of Music and Arts showed that the average performance of the students in the experimental class that utilized the system for assisted teaching increased by 9.00 points (from 65.45 to 74.45), while the control class that did not utilize the system only increased by 0.79 points, which is a significant difference (P=0.041<0.05). In addition, teaching based on the piano hand fingering recognition system significantly increased students' interest in learning, with the experimental class increasing its interest score from 38.54 to 48.49. The study confirms that the deep learning-based piano hand shape fingering recognition system can effectively improve students' piano playing skills, stimulate learning interest, and provide a new teaching aid for piano art instruction.
