Fusion of superpixel, expectation maximization and PHOG for recognizing cucumber diseases

- Propose a fast algorithm for lesion segmentation from the diseased leaves.
- Propose a new feature representation for cucumber disease lesions.
- Proposed method outperform comparison methods in both recognition rates and computation cost.

Cucumber diseases can be detected and recognized automatically based on diseased leaf symptoms. In this paper, we propose a new method, combining superpixels, expectation maximization (EM) algorithm, and logarithmic frequency pyramid of histograms of orientation gradients (PHOG), to recognize cucumber diseases. The proposed method is composed of following steps. First, the superpixel operation is used to divide a diseased leaf image into a number of compact regions, which can dramatically accelerate the convergence speed of the EM algorithm that is adopted to segment the diseased leaf regions and obtain the lesion image. Second, the logarithmic frequency PHOG features are extracted from the segmented lesion image. Finally, Support Vector Machines (SVMs) are performed to classify and recognize different cucumber diseases. Conducted on a database of cucumber diseased leaf images, experimental results show the proposed method is effective and feasible for recognizing cucumber diseases.