Abstract
Slight cracks in cottonseed significantly impact germination rates by allowing water or acid to penetrate the seed coat during processing. Because these defects have minimal impact on seed morphology or chemical composition, traditional non-destructive methods like machine vision and optical spectroscopy often fail. This study introduces a novel detection method using air-coupled ultrasound combined with a sound-to-image encoding technique and a lightweight Vision Transformer (MobileViT). By converting ultrasonic echo signals into colorful 2D images, the system achieves a 90.7% recognition accuracy, offering a reliable, non-contact solution for seed quality assessment.
The Challenge of "Invisible" Defects
In the cotton industry, seed quality is paramount for ensuring high crop yields. During the ginning and stripping processes, cottonseeds frequently develop "slight cracks"—tiny fissures in the testa (seed coat). While these cracks are often microscopic, they are detrimental; they allow moisture or processing chemicals to seep into the seed, drastically reducing its viability.
Detecting these cracks is notoriously difficult. Standard non-destructive techniques face specific limitations:
- Machine Vision: Cracks are often too small to be resolved or are located on the side of the seed facing away from the camera.
- Optical Spectroscopy: Since the chemical makeup of the seed remains unchanged, the spectral signature does not shift.
- Thermal Imaging: There is no significant temperature difference between intact and slightly cracked seeds.
Acoustic methods, however, are highly sensitive to structural integrity. The challenge lies in implementing ultrasound in a way that is both non-contact and high-throughput, avoiding the contamination associated with traditional liquid couplants.
A Novel Multi-Stage Detection Pipeline
The researchers developed a pipeline that transforms 1D acoustic signals into a format suitable for advanced computer vision models.
1. Air-Coupled Ultrasound
The system utilizes a pair of 400 kHz transducers to transmit and receive ultrasonic waves through the seeds. By using air as the coupling medium, the process remains entirely non-contact and non-destructive.
2. Signal Decomposition and Encoding
Ultrasonic echoes are complex and non-stationary. To extract meaningful features, the team employed Variational Mode Decomposition (VMD) to break down the 1D signals into Intrinsic Mode Functions (IMFs).
The most innovative step is the Sound-to-Image Encoding. Instead of processing raw numerical data, the IMFs are converted into colorful 2D polyline images. By using different colors to represent changes in IMF values over time, the researchers transformed a signal processing task into a pattern recognition problem. Experiments revealed that a 10-color encoding scheme provided the optimal balance of information density and feature clarity.
3. Lightweight Deep Learning
The encoded images are processed by MobileViT, a lightweight Vision Transformer. MobileViT is uniquely suited for this task as it combines the local feature extraction capabilities of Convolutional Neural Networks (CNNs) with the global attention mechanisms of Transformers. This hybrid architecture allows the model to understand both the fine "textures" of the signal and its overall structural patterns.
Results and Key Insights
The study evaluated 600 cottonseeds (296 intact and 304 with slight cracks). The proposed VMD-CE-MobileViT method demonstrated superior performance:
- Accuracy: 90.7%, significantly higher than traditional 1D-LSTM models (74%).
- Reliability: The system achieved a recall of 96.4%, ensuring that very few damaged seeds are missed.
- Safety: Germination tests on 2,000 seeds confirmed that the 400 kHz ultrasound does not negatively affect seed viability.
The success of this approach highlights the power of sound-to-image encoding. By providing spatial context to acoustic data, the model can "see" relationships between frequency components that are difficult to capture through traditional mathematical analysis.
Future Directions
While the current results are promising, future work will focus on testing the system across different cottonseed varieties with varying shell thicknesses. Additionally, the researchers aim to optimize the system for industrial-scale throughput. The "sound-to-image" encoding technique also holds potential for other fields, such as the analysis of medical signals like ECGs.
Summary updated