PA2E: Real-Time Anomaly Detection With Hyperspectral Imaging for Food Safety Inspection

Jungi Lee,Myounghwan Kim,Jiseong Yoon,Kwangsun Yoo,Seok-Joo Byun

DOI: https://doi.org/10.1109/access.2024.3505147

IF: 3.9

2024-12-04

IEEE Access

Abstract:Hyperspectral imaging captures material-specific spectral data, making it effective for detecting contaminants in food that are challenging to identify using conventional methods. In the food industry, the occurrence of unknown contaminants is problematic due to the difficulty in obtaining training data, highlighting the need for anomaly detection algorithms that can identify previously unseen contaminants by learning from normal data. Deep learning techniques for anomaly detection have been extended to hyperspectral imaging. However, existing methods often use fully connected layers that do not leverage local spectral similarities or convolutional layers that overlook positional importance due to transform invariance, leading to performance limitations. Furthermore, practical hyperspectral contaminant detection in food requires significant real-time processing capabilities, hindered by the computational inefficiency of convolution operations and the limitations of existing deep neural network architectures. To address these issues, we propose a novel model called Partial and Aggregate Autoencoder (PA2E), which incorporates local information without transform invariance and utilizes masked fully connected layers to achieve high inference speed. Additionally, optimizations such as ReLU culling and layer fusion were applied to the fully connected architecture, dramatically reducing inference time without compromising accuracy. Validation on three hyperspectral datasets demonstrated that PA2E achieved an average detection capability of 29, outperforming the current state-of-the-art methods (average of 23 out of 42) with 2.6 times faster inference speed. PA2E leverages the unique characteristics of hyperspectral data and offers a highly optimized layer structure, making it a promising model for application in the hyperspectral food industry and other hyperspectral domains.

computer science, information systems,telecommunications,engineering, electrical & electronic