Related Experiment Video
Updated: Jun 30, 2025
06:41
Imaging and Analysis for Quantifying Maize (Zea mays) Abiotic Stress Phenotypes
Published on: March 28, 2025
799
Non-destructive detection of defective maize kernels using hyperspectral imaging and convolutional neural network
Dong Yang1, Yuxing Zhou1, Yu Jie1
1Academy of National Food and Strategic Reserves Administration, Beijing 100037, China; National Engineering Research Center of Grain Storage and Logistics, Beijing 100037, China.
Summary
This study introduces a hyperspectral imaging (HSI) method with a convolutional neural network (CNN) featuring spectral and spatial attention for detecting defective maize kernels. The advanced CNN model achieved high accuracy in identifying various kernel defects non-destructively.
Area of Science:
- Agricultural Engineering
- Computer Vision
- Spectroscopy
Background:
- Accurate detection of defective maize kernels is essential for maintaining grain quality during storage.
- Traditional methods for defect detection are often time-consuming, destructive, or lack precision.
- Hyperspectral imaging (HSI) offers a promising non-destructive approach for analyzing agricultural products.
Purpose of the Study:
- To develop and evaluate a hyperspectral imaging (HSI) system combined with a convolutional neural network (CNN) for the non-destructive identification of defective maize kernels.
- To investigate the effectiveness of spectral and spatial attention mechanisms within CNN models for improved kernel defect recognition.
- To compare the performance of the proposed attention-based CNN models against conventional machine learning algorithms.
Main Methods:
- Collected hyperspectral imaging (HSI) data (380-1000 nm) for six classes of maize kernels: sprouted, heat-damaged, insect-damaged, moldy, broken, and healthy.
- Developed and trained three CNN models: CNN-Spl-At (spectral attention), CNN-Spal-At (spatial attention), and CNN-Spl-Spal-At (fused spectral-spatial attention).
- Compared the CNN models with Support Vector Machine (SVM) and Extreme Learning Machine (ELM) using spectral, image, and fused features.
Main Results:
- CNN models with attention mechanisms significantly outperformed SVM and ELM models in kernel defect recognition.
- Fused spectral and spatial features provided more comprehensive information for distinguishing kernel types than single features.
- The CNN-Spl-Spal-At model achieved the highest accuracy, with 98.04% for the training set and 94.56% for the testing set.
Conclusions:
- The proposed CNN-Spl-Spal-At model, utilizing HSI and attention mechanisms, effectively detects various defective maize kernels non-destructively.
- This approach demonstrates significant potential for developing advanced, non-destructive testing equipment for real-time maize quality assessment.
- The integration of spectral and spatial information via attention modules is crucial for enhancing classification accuracy in HSI applications.
