Towards Content-Based Image Search
In this project, an autoencoder neural network was developed and the encoder section was used to generate the embedded representation of the images within the database. An image was considered similar to a selected image, if they were close to one another in the embedded space. The search experiments were further extended to accommodate two image inputs in which a variational autoencoder was used to generate an image based on two input images, and the generated image was used to search for relevant images in the database using the described search method. The validity of this approach was tested on the ORL and CIFAR-10 databases and yielded an accuracy of 33% to 48% in different cases.