Abstract: An improved one-dimensional convolutional neural network (1D-CNN) model for the authenticity detection of egg white powder was constructed based on near-infrared spectroscopy (NIRS). This model required no spectral preprocessing. To enhance its ability to extract spectral features, an efficient channel attention module (ECA) and a one-dimensional global average pooling (1D-GAP) layer were singly or together incorporated into the model, consequently reducing noise interference. The experimental results indicated that the improved model integrating ECA and 1D-GAP, EG-1D-CNN, could distinguish between authentic and adulterated egg white powder samples, with a detection rate of 97.80% for adulterated samples and an overall accuracy rate (AAR) of 98.93%. The lowest recognition concentrations (LLRC) for single adulterants such as starch, soy protein isolate, melamine, urea, and glycine were 1%, 5%, 0.1%, 1%, and 5%, respectively, and those for multiple adulterants ranged from 0.1% to 1%. The average time spent (AATS) for the detection was 0.004 4 seconds. Compared with traditional 1D-CNN network structure and other improved algorithms, the EG-1D-CNN model exhibited higher accuracy, faster detection speed, and smaller model footprint, thus making it more suitable for deployment on embedded devices. This research provides a theoretical foundation for the development of portable near-infrared spectroscopy-based detectors for egg powder quality testing.

Key words: egg white powder; near-infrared spectroscopy; authenticity detection; one-dimensional convolutional neural networks; deep learning