Dance movement analysis requires accurate image segmentation, action recognition, and movement quality assessment. However, in traditional systems, these operations are carried out independently, and no relationship is established between the accuracy of segmentation and the results of classification. This study proposes an end-to-end framework integrating an ADE-GJO hybrid optimization-based multi-threshold image segmentation method, a segmentation-guided spatio-temporal attention action recognition network termed SG-STANet, and a dynamic time warping-based movement deviation measurement mechanism. The ADE-GJO algorithm incorporates chaotic initialization, adaptive DE-based mutation operation, and Lévy flight in the GJO algorithm to enhance its global exploration capability at high threshold levels. The SG-STANet model uses foreground information provided through segmentation as a prior for spatial attention to prevent any negative effect of the background on skeleton feature extraction. Experiments performed on BSDS500, AIST++, and NTU RGB+D 120 show that the introduced segmentation algorithm achieves decent results. Mask-guided mechanism results in significant improvements for classification purposes, while the DTW similarity measure successfully recognizes different dance styles based on their quality by means of native data set annotations. Overall, the introduced approach provides a technically reasonable pathway toward the use of computer vision technologies in evaluating dance education and intelligent training.
