TECHNIQUE FOR NORMALIZING INTENSITY HISTOGRAMS OF IMAGES WHEN THE APPROXIMATE SIZE OF THE TARGET IS KNOWN: DETECTION OF FECES ON APPLES USING FLUORESCENCE IMAGING

Authors: Lefcourt, Alan; Kim, Moon

Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 10/6/2005
Publication Date: 2/20/2006
Citation: Lefcourt, A.M. and Kim, M.S. 2006. Technique for normalizing intensity histograms of images when the approximate size of the target is known: Detection of feces on apples using fluorescence imaging. Computers and Electonics in Agriculture. 50:135-147.

Interpretive Summary: Imaging is gaining importance as a technique for automatically examining agricutural products. One area of particular interest is detecting contamination on food products. Often the contaminate has no particular shape, e.g. any contamination due to splashing, so techniques for recognizing shapes are not useful. Detecting the contaminate is even more difficult when measured responses for both the contaminate and the food product are due to substances found in both. Detection is even more difficult when the amount of these substances in individual units varies. An example of this problem is detection of fecal contamination on apples. Feces on apples can be detected using fluorescence responses to ultraviolet light. However, responses for both feces and apples are due to the presence of chlorophyll-related compounds, and the response of apples varies within and between apples due to natural variation in the distribution of these compounds. Presented is a technique for normalizing variable intensity responses among apples based on knowledge of the image dimensions and the general size of apples. Using this information, a linear equation is derived based-on the measured intensities of the background around the apple and of the apple. The image is scaled using this equation. The benefits of using this technique are demonstrated using 48 Golden Delicious and 48 Red Delicious apples that were artificially contaminated with cow feces. Results show that by using these derived-equations, the intensity of apples is consistent from apple to apple, and the contrast between feces and apples is increased. This technique can be used to improve the ability to detect some contaminates in both scientific and commercial applications.

Technical Abstract: For machine vision, it is a challenge to develop algorithms for detecting a substance with an amorphous shape when measured responses of both the substance and the background target are similar. The challenge is exacerbated when responses for targets are highly variable both across and within discrete target units. An example of this problem situation is the detection of fecal contamination on apples. Feces on apples can be detected using differential fluorescence responses to UV excitation. However, responses for both feces and apples are due to the presence of chlorophyll-related compounds, and the response of apples varies within and between apples due to natural variation in the distribution of these compounds. Presented is a technique for normalizing variable intensity responses among targets based on a priori knowledge of the image dimensions and the approximate target size. Using this information, a linear equation is derived based-on the median intensities of the background and the target. The image is scaled for uniform intensity power using this linear transform. The benefits of using this technique are demonstrated using hyperspectral fluorescence responses to UV excitation of 48 Golden Delicious and 48 Red Delicious apples that were artificially contaminated with dilutions of cow feces. Results show that the uniform power transform normalizes the intensity distributions of apple images and increases the contrast between contaminated and untreated areas on apple surfaces.