Sub-Category Classification in Produce Recognition Systems: Evaluating Embedding-Based Classifiers in Euclidean and Spherical Space

The purpose of this thesis is to discuss research conducted during a six-month internship at Datalogic USA, with a focus on the optimization of the second-level classifier in a hierarchical automatic produce recognition system. The existing system relies on a DenseNet backbone network for the generation of embeddings, which are classified using class-specific Gaussian Mixture Models for the final classification. However, one of the most important characteristics of the second-level classifier is the support of on-premise incremental learning, which allows individual locations of the stores to update the product catalogs without having to retrain the entire network. This research can be classified into two major phases. First, the evaluate the reliability and robustness of the current GMM method in comparison with other machine learning techniques. These techniques include K-Nearest Neighbors, Random Forest, Support Vector Machines, and Logistic Regression etc.. Although some of these methods, like Logistic Regression or Gaussian Naive Bayes, have shown superior results in terms of accuracy, they have not shown sufficient modularity, separability, class independence or extensibility required in the production environment. Hence, the probabilistic mixture models remain the first choice. The second phase of the research process focuses on the geometric mismatch problem with the implementation of hyperspherical embedding techniques. As the research process evolved with the implementation of ArcFace and AdaCos loss functions, these functions are used to optimize the angular margins rather than the Euclidean distances. In that case the Gaussian assumption become theoretically suboptimal in a non-euclidean space. To overcome this problem, a novel classifier called von Mises-Fisher Mixture Model (vMFmm) was implemented. It models the distribution of the data on the hypersphere surface, which acts as a spherical equivalent to the GMM.