A multiclass neural network model for contaminant detection in hazelnut-cocoa spread jars

This project aims at developing a machine learning algorithm for detecting six different types of contaminants in hazelnut-cocoa spread jars. In particular, work began by analyzing an already existing binary, 3 layer deep learning classifier (DNN) which presented a misprediction rate of 51% for a triangle shaped plastic. New DNN architectures were developed, trained and validated using a multiclass approach instead after careful examination of the available samples and previous preprocessing methods. After optimization procedures were finished using Grid Search and Stratified Cross Validation for initial diagnostics, the best candidate turned out to be a 1 layer MLP architecture. This model managed to reduce the error rate for the problematic contaminant from 51% to 7.5% in the worst case, This was calculated after analyzing all relevant confusion matrices during the testing phase and reduced to a 2x2 dimension to compare results to the binary case. Finally, the model was mapped into hardware and uploaded to an FPGA using the ZCU106 board with the software.