Cassava disease detection by fractal analysis of digital images

A crop monitoring system that provides early detection of plant disease is indispensable for future crop protection. Diseased plants are usually observable from visual symptoms and hence suggest that possibility of damaged plants identification using machine vision (L). This requires a quantity which is capable of characterizing atypical appearances of the diseased plant organs. This study evaluated the validity of some fractal parameters in plant disease identification. Colored images of normal and diseased cassava leaves were captured with a digital camera at a horizontal by vertical resolution of 1600 by 1200 pixels. The original color images were grayscaled and thresholded to detect their edge outlines. These were converted to binary images prior to analysis. Fractal analyses based on box-counting algorithm were performed for individual image to estimate fractal dimensions and lacunarities using Fraclac v. 2.49 for Image J. The software functioned in placing non-overlapping grids over the image by varying the grid size from 2 to 540 pixels. The number of boxes required to cover foreground pixels as well as the number of pixels in each box were counted. Two types of fractal dimensions i.e., standard box-count dimension (Dstd), and minimum cover box-count dimension (Dmc) were calculated. The slope corrected Dstd and Dmc which eliminate the effect of plateaus slope were also observed. The lacunarities considered includes prefactor lacunarity (lambda pf) derived both from box counts and pixel masses, overage lacunarity (lambda avg) and binned probability density lacunarity (lambda bpd) obtained from foreground masses only, and foreground plus empty boxes. The results showed that Dstd and Dmc of diseased leaves images were significantly greater than those of normal leaves which were similar to sloped corrected Dstd and Dmc. This indicates that the diseased regions appeared on the cassava leaves had introduced a marked complexity to the leaves profiles. The lacunarities of diseased leaves were found smaller than those of typical leaves for all cases. Nevertheless, these differences were not statistically significant except for the lambda avg obtained from foreground plus empty boxes. This implies that alteration in spatial heterogenity due to the diseased regions, under the present condition, was not obvious. The fractal dimensions are therefore of promising parameters make the diseased plant identifiable.