A three-dimensional threshold algorithm based on histogram reconstruction and dimensionality reduction for registering cucumber powdery mildew

Abstract

© 2019 Elsevier B.V. Analysis of plant disease images can increase disease detection accuracy. Successful extraction of lesions can provide a good precondition for research on the feature analysis (texture, color and shape) and detection of occurrence area and severity for cucumber powdery mildew. This paper proposes a new Otsu algorithm based on three-dimensional histogram reconstruction and dimension reduction to segment powdery mildew from cucumber disease images. In total, 166 RGB images of cucumber powdery mildew with 1440 × 1080 pixels were obtained from the greenhouse nos. 1 and 2 using a high-speed dome camera. First, a new correction formula is proposed to correct anomalous points to the correct position (around the line connecting the origin with diagonal end). Second, we reduced the algorithm dimension to save time and spatial complexity and achieved the desired results. Finally, the optimal threshold was obtained by a Gaussian fitting iteration. Convergence analysis demonstrated that the method should be able to obtain a new threshold after each iteration. Experimental results showed an average false negative error rate of 0.10% and average false positive error rate of 1.27%. The average running time was 5.082 s. Taken together, these represent satisfactory results for rapid automatic identification of cucumber powdery mildew.