Dance forms play a vital role in displaying the cultural heritage of a country. The costumes play a vital role not just in creating a visual impact and elevating the performance but also defines the dance form itself. Dance costume identification is a quite complex task as the costume worn by the dancers look similar while different dance forms are compared. In this paper, we propose a novel method that can be used for classifying four different Indian classical dance forms, by considering features of the costume alone. A novel dataset has been built and two different preprocessing methods were adopted—Background removal on RGB images and Canny edge detection algorithm applied over background removed images. A CNN and a pre-trained VGG 16 model were used for the extraction of costume-based features from the images. Training accuracy of 99.92%, validation accuracy of 73.84% and testing accuracy of 82.09% were obtained for the CNN model and with VGG 16, training accuracy of 95.92%, validation accuracy of 92.92% and testing accuracy of 82.21% were obtained. The results are proven to be better when background removal on RGB images is used as preprocessing technique with VGG 16 model adopted for training. Costume-based identification enhances dance recognition by improving accuracy, especially in crowded or visually complex environments, by providing unique visual cues for individual dancers and will be of much use to the researchers who wish to explore the intricacies in the field of dance.