Naïve Bayes pixel-level plant segmentation

In this paper, we propose a naïve Bayes classification method for pixel-level plant segmentation, which can classify plant and background (plant/non-plant) with high accuracy in an image. Unlike existing algorithms, in which statistical information is extracted from a single image and optimization techniques are applied to minimize the segmentation cost function, in the proposed method, statistical information/ features are extracted from a training image dataset (each image includes plant and non-plant background) and a naïve Bayes classification method is applied to identify each pixel class in an image. Therefore, the proposed method is less complex than other learning techniques as no (regularized) optimization method need to be applied in the segmentation algorithm. The proposed method is more useful when a limited number of classes/objects needs to be classified, such as the two-class plant/non-plant problem. An image dataset with sufficient images to extract feature information in the training step is also needed. The proposed method is a suboptimal classification but can approach the optimum one if the pixels probability distribution functions (PDFs) are approximated accurately and a posteriori probability of each pixel is available.