Identification of immature tomatoes base on salient region detection and improved Hough transform method

Abstract: The identification of fruit crop image plays an important role in the automatical estimation of production. However, occlusion, varying illumination, and similarity with the background make fruit identification become a very challenging task. Green tomato detection with green canopy is a very difficult problem. In this paper, we first put forward the dense and sparse reconstruction (DSR) method to detect immature fruit images. This method first computes dense and sparse reconstruction errors. Second, the reconstruction errors are propagated based on the contexts obtained from K-means clustering. Third, pixel-level saliency is computed by an integration of multi-scale reconstruction errors and refined by an object-biased Gaussian model. Finally this method applies the Bayes formula to integrate saliency measures based on dense and sparse reconstruction errors. The DSR detection images of tomato are obtained by the computation mentioned previously. Then the OTSU method is used to segment the DSR detection images, and the opening operation is used for removing the noise area after the split. In particular, in order to identify single fruit out of tomato fruit clusters, the circular Hough transformation (CHT) or randomized Hough transform (RHT) can be used, but they have some shortages (e.g. a large amount of calculation and space overhead) when recognizing tomato fruit. So we trade off the advantages and disadvantages of CHT and RHT, and propose an improved randomized Hough transform (IRHT) circle detection method. First, we adopt the boundary tracking algorithm to obtain tomato image boundary after segmentation. Secondly, in order to obtain circles and radii, we utilize the subsection and interval point group selection method to improve the accuracy of identification and reduce redundant computation. However, some of circles and radius may be in the same circle, or may be too big or small, so we need to find the real circles. Finally, in order to eliminate repetitive circles to get the last circle target (the fruit), we compute the Euclidean distance of center coordinates of 2 circles, and accumulate those circles who are in the same real circle to generate actual circle (a tomato fruit). In this paper, we compare the test results of 3 methods, namely the traditional CHT, RHT method and our proposed method. And we find that the correct rate of our proposed improved algorithm of immature image recognition can reach 77.6%. Moreover, the correlation coefficients of circle centers and radii of tomato fruit between ground truth and calculated value by our method are 0.98 and 0.76, respectively. The average relative error of center coordinates of circle is 7.6%, and the average relative error of radii of circle is 14.0%. The confidence level of mean in the confidence interval (42.03, 49.48) is 95%, and the confidence level of variance in confidence interval (10.25, 15.64) is 95%. Based on the results of our study, we find that the universal applicability of our method is stronger, in addition, our method is not only suitable for tomatoes, but also applicable to other kinds of cone crops. Therefore, our method lays a solid foundation to achieve the goal of production estimates of a variety of fruits by robots.