Computer-vision classification of corn seed varieties using deep convolutional neural network

TLDR: A new approach using a deep convolutional neural network (CNN) as a generic feature extractor for intelligent classification of different corn seed varieties is presented.

About: This article is published in Journal of Stored Products Research.The article was published on 2021-03-20 and is currently open access. It has received 48 citations till now. The article focuses on the topics: Convolutional neural network & Feature (machine learning).

Figures

Table 1 Extracted morphological features of corn seeds (Medina et al., 2010; Zhao-Yan et al., 2005).

Fig. 2. Results according to the image preprocessing and segmentation processes.

Table 2 Texture features extracted from GLCM matrix (Pourreza et al., 2012).

Fig. 7. The confusion matrices of the different classifiers on the testing data set based on the features that had the best performance as described in Table 4: (a) CNN þ color; (b) CNN þ color; and (c) CNN þ LBP. Here 1, 2, 3, 4, 5, 6, 7, 8, and 9 are depicting KSC 201, KSC 704, KSC 290, KSC 380, KSC 301, KSC 400, KSC 260, KSC 647, and KSC 410.

Fig. 3. The overall structure of the VGG-16 for feature extraction. The figure also shows the visualization of the features of different blocks.

Table 3 Specific parameters of the CNN architecture.

Frequently asked questions

Q1. What are the activation functions used in a CNN?

There are several activation functions commonly used in CNNs, such as the Rectified Linear Unit (ReLU), sigmoid, and hyperbolic tangent function. 

Q2. What are the contributions mentioned in the paper "Computer-vision classification of corn seed varieties using deep convolutional neural network" ?

This paper presents a new approach using a deep convolutional neural network ( CNN ) as a generic feature extractor. 8 s with classification accuracy 98. 1 %, precision 98. 2 %, recall 98. 1 %, and F1-score 98. 1 %. This study demonstrates that the CNN-ANN classifier is an efficient tool for the intelligent classification of different corn seed varieties. 

Q3. What is the common term used to describe the average error between calculated output and target output?

Cross-entropy is commonly used to describe the average error between calculated output and target output in the logarithmic scale. 

Q4. What are the standard analytical methods used to classify plant varieties?

High-performance liquid chromatography, gas chromatography-mass spectrometer (Qiu et al., 2018), seed protein electrophoresis (Rogl and Javornik, 1996), and DNA molecular markers (Hoffman et al., 2003) are some of the standard analytical methods used to classify plant varieties. 

Q5. What was the classification time for the bagged tree classifier?

When individual hand-crafted features were used to classify the classes, a weighted kNN classifier based on color features required the shortest classification time among all algorithms. 

Q6. How does the SVM model separate the training vectors?

Instead of modeling the probability distribution of training vectors, SVM attempts to separate them by directly searching appropriate boundaries between different classes. 

Q7. How many seeds were stored in sealed plastic packages?

In the healthy seeds set, 1000 samples from each variety were randomly selected for imaging and stored in sealed plastic packages at room temperature (20 ± 1 C). 

Q8. How many samples were misclassified as other classes?

In class 3, nine samples were classified wrongly as other classes and in class 6, there were 11 samples misclassified as other classes. 

Q9. Why did the authors present cross-entropy performance and error histogram only for this case?

Because the CNN-ANN configuration had the best performance and accuracy, the authors presented cross-entropy performance and error histogram only for this case (Figs. 4 and 5). 

Q10. Why was the average accuracy of classification lower than 45%?

The average accuracy of classification based on morphological features was lower than 45% because of the high similarity between morphological features of corn varieties. 

Q11. How many misclassifications were found in the cubic SVM model?

An obvious conclusion could be drawn from Fig. 6(b) that in the cubic SVM model based on the fusion of CNN and LBP features it was 99 misclassifications, while for the ANN model trained with CNN features only 42 corn seeds were misclassified. 

Q12. How many iterations of the validation set were used?

As visible in Fig. 4, the best validation performance is obtained at a minimum cross-entropy error of 0.0028687 in 165 iterations.