Recognizing black point in wheat kernels and determining its extent using multidimensional feature extraction and a naive Bayes classifier

Accurately identifying black point disease in wheat kernels from random samples within digital images is a fundamental and challenging task in disease identification. The performance of traditional methods is satisfactory in homogeneous environments, but their performance decreases when they are applied to images acquired in dynamic ones. In this paper, a multifeature-based machine learning method is proposed to identify and evaluate the incidence of black point disease. Ten wheat cultivars with different resistances to disease were selected to verify the accuracy of the method. First, a marker-based watershed algorithm was used to separate wheat kernels from the background to accomplish the coarse segmentation. After patches were generated from the coarse segmentation results, the patches were labeled manually and divided into two categories: black point areas and healthy areas. Gabor and Canny operators were used for texture and shape features respectively to build a feature matrix. Then, a classification model based on a naive Bayes classifier was trained to recognize the differences between the two types of patches by their features. The proposed model finally achieved the correct classification of each pixel from the testing sample and output the results in the form of a binary image, thus accomplishing the segmentation of the image. Finally, the severity of the disease was calculated according to the proportion of minimum circumscribed areas of the disease and the total area of the wheat kernel. Through the above operations, the incidence of black point disease in random samples can be determined. Five indicators, Qseg, Sr, Precision, Recall, and F-measure, were used to evaluate the segmentation effects. The average accuracy of segmentation results for the testing samples were 0.85, 0.89, 0.87, 0.86, and 83% respectively. Compared with other segmentation approaches, including the excess green method, the excess green minus excess red method, the color index of vegetation extraction, and two traditional threshold segmentation methods known as Otsu and maximum entropy threshold, the proposed algorithm had greater segmentation accuracy. Moreover, this method was demonstrated to be robust enough to be used for different illumination conditions, shooting angles, and image resolutions.