Gırtlak mikrofonu kayıtları üzerinden öğrenim aktarımının otomatik diyet takibi için kullanımı

Wearable devices and technologies in healthcare have been accelerating the development and integration of engineering approaches. Dietary monitoring is one challenging application among other healthcare services. Personal records are typically preferred ways of dietary monitoring. However, manual logging is highly biased and unreliable as individuals tend to underestimate their food intake. Automatic dietary monitoring (ADM) is an intelligent wearable coaching solution for this problem. The current research on ADM is shaped by the sensing devices and is fitted by machine learning approaches. In this thesis, we first define an ADM system using a throat microphone (TM) food intake sound recordings, where chewing and swallowing detection is presented. This system uses the TM sensor as a non-invasive transducer mounted on the neck, which is capable of delivering robust signal recordings for intelligent and unobtrusive food intake monitoring. Then, we investigate the use of transfer learning paradigm in depth to design an improved food intake detection and classification system. Although it is possible to reach abundant labeled food intake data recorded with traditional close-talk microphones (CM), data from TM modality is scarce. This creates a bottleneck in training deep architectures effectively using TM data. Recently, teacher/student (T/S) learning paradigm is introduced as a model compression framework, where it describes a class of learning methods for training a smaller student network by mimicking a larger teacher network. We propose a new domain adaptation framework in a heterogeneous setup based on T/S learning paradigm. The teacher network is trained over abundant high-quality CM recordings, whereas the student network takes TM recordings as input and distills deep feature extraction capacity of the teacher over a parallel CM and TM dataset. This allows the use of a significantly larger set of adaptation data, adds robustness to the resulting model, and significantly improves the performance of food intake detection.