Food texture is a key determinant of palatability. While human sensory testing remains crucial for replicating the final consumer experience, obtaining reliable results remains challenging even with established methods like TI, TDS, and TCATA. A significant challenge lies in the discrepancy between TPA characteristic values and sensory test results, probably due to variations in experimental conditions such as tooth morphology, temperature, saliva composition, flavor crosstalk, and masticatory patterns (e.g., chopping and gliding).

Deep learning has gained considerable attention due to its potential to categorize data even with subtle differences. Even for above mentioned type of complex task of food texture, deep learning may be applicable. Huge size of dataset is crucial in general for deep learning, depending on model complexity. Traditional food science research has been limited by the small datasets available (typically on the order of 10~100 measurements). This limitation has restricted the application of deep learning to relatively simple analyses to prevent overfitting.

We have developed automated systems for collecting food compression data and have collected a dataset exceeding 10 ⁵ measurements using conventional 1-axial and 6-axisal force/torque sensors with teeth shaped plunger. This dataset allows for the detection and analysis of subtle differences, even in foods exhibiting significant textural variations where TPA cannot be applied. The large dataset demonstrates that MLP has the potential to analyze food texture using a fundamentally different approach compared with TPA.

Despite the extensive dataset of compression measurements, a critical gap still remains in replicating the variability of experimental conditions encountered during instrumental compression measurements. To address this gap, 3D scanning of facial movements during chewing using a smartphone-integrated 3D sensor was employed to collect a large-scale dataset of human mastication behavior. Chewing behavior is significantly influenced by both food texture and flavor profiles. A dataset exceeding 10 ⁵ of mastication behavior has already been collected. More detailed analysis tightly related to sensory evaluation was possible with mastication behavior compared with compression measurements.

Deep learning-based food texture analysis integrating compression and chewing measurements provides a promising way for detecting subtle textural differences and bridging the gap between human sensory scores and instrumental evaluations.