Bayesian optimization of convolutional neural network for identifying red wines using colorimetric sensors

Integrated low-cost sensors with deep learning frameworks exhibit excellent discrimination capabilities across various applications. However, several factors influence the performance of these systems, making robust calibration and adaptive learning strategies essential for maintaining reliability under varying conditions. A low-cost, deep learning (DL)-assisted colorimetric sensor (CS) system was developed in this study to identify nearly identical red wines efficiently. Commercial red wine samples were tested using CSs coated with chemically responsive dyes, and images of the sensors were captured using a smartphone camera before and after exposure to the wine samples. An automatic feature extraction and multiclass classification framework is employed to identify wine samples directly from the acquired image data, rather than from handcrafted numerical features derived from the absolute RGB differences between pre- and post-exposure images. Handcrafted features generated by image processing software, such as ImageJ, are typically used to interpret CS data. However, this process can be time-consuming, prone to errors, and yield lower accuracy, particularly with complex datasets involving multiple classes. A Bayesian optimization framework was used to determine optimal hyperparameters for a convolutional neural network (CNN), achieving classification accuracy of approximately 95 %. For comparison, handcrafted features were also analyzed using classical machine learning algorithms. The results showed that taste identification using an automatic feature extraction framework is an effective method for discriminating sensor data. This study demonstrates an efficient method for identifying red wines using low-cost sensing integrated with an optimized DL model.