Image segmentation for whole tomato plant recognition at night

Abstract

A simplified pulse-coupled neural network was developed based on comentropy gradient for segmentation of tomato plant images captured at night. In this network, the connection matrix was determined using differences in the value and distance between the center and other pixels in a neighborhood. To minimize the number of iterations, the model used linear threshold attenuation function on a limited range. A new evaluation method based on comentropy gradient was also used to determine the best image segmentation result during iteration. Evaluation of 385 tomato plant images captured at night showed that the simplified pulse-coupled neural network segmented the images of tomato plants captured at night. The method improved the image segmentation performance, with best and false rates of 67.79% and 9.61%, respectively, for all segmentation results. In addition, the best and false rates are 59.22% and 13.77%, respectively, for the method based on the maximum comentropy. The proposed algorithm exhibited the optimal segmentation performance, with the best rate of 91.67%, the second best rate of 8.33%, and the false rate of 0.00% for images with 640 × 480 pixel captured at distances from 300 mm to 500 mm under various lighting conditions by using the lighting system composed of two 25 W incandescent light bulbs in the opposite layout at night. The linear threshold attenuation function improved the real-time performance of the method. For 640 × 480 and 320 × 240 pixel images, the average running times were 2.69 and 0.61 s, respectively, compared with the original times of 12.02 and 2.90 s.