USE OF SIAMESE CONVOLUTIONAL NEURAL NETWORKS TO COMPUTE IMAGE SIMILARITY

This work presents the usage of siamese convolutional neural networks as a solution for computing image similarity/distance between two colored images. Itsmaincontributionscan be divided in three parts. The practical applications are contextualized inabusinessscenario related to pricing strategy. Regarding the main contributions, first, the training ofdifferentsiamesemodelsispresented, along with their results using the public dataset ​DeepFashion as training and test data. An offline method for constructing a labeled data set of similar and non-similar image pairs is presented. Such data set is then used to train and evaluate the different siamese models.The models presented can differbothintheselectionoftheCNNbranchesandintherankingloss function used for training (metric learning). The final results include other approaches for computing image similarity as well, such as Mean Squared Error and Structural Similarity. The second maincontributionconsistsofanextensionoftheclassiccontrastivelossfunction. Such function is usually applied with a single margin, however there exists literature of an application using a double margin contrastive loss. In this work, both losses are tested with the objective of determining whichoneproducemodelsthatperformbetterintermsofimage ranking. A generalization of an N-margin contrastive loss is introduced as an idea, as a motivation for future work. The third and final part presents the results of a real business case application. Such application uses both image and text information in order to match products that are considered similar, to then compareitspriceswiththeobjectiveofknowinghowcompetitive are the prices of a given company.