Deep Learning Approach for Expression-Based Songs Recommendation System

Both music fans and consumers may appreciate music, which is a great method for individuals to express themselves. With the development of technology, there are more musicians and more people who enjoy music, and this raises the challenge of manually selecting music. Song suggestions have been around for a while, but in most cases, they are created after learning about the user’s tastes over time. For example, they may take into account the user’s prior song selections, how often he listens to music, and other factors. In this research, we suggest a novel method of song suggestion in which a person’s mood is predicted from his or her photo, and songs are then suggested that most closely match the mood indicated. FER 2013 was used to train this model. Images of faces in greyscale measuring 48 × 48 pixels make up the data. As a result of the faces being automatically registered, each picture has a face that is about in the centre and takes up nearly the same amount of area. Each face must be assigned to one of seven categories, with zero denoting anger, one disgust, two fear, three happiness, four sadness, five surprise, and six neutral. A total of 3589 examples make up the public test set, whereas 28,709 instances make up the training set. We are employing Resnet50 in place of VGG16 and normal sequential CNN. After building the model, we are fine-tuning it again to get more accurate results.