Optimal Selection and Identification of Defects in Chestnuts Processing, through Computer Vision, Taking Advantage of its Inherent Characteristics

Abstract

In the agro-industry automation, computer vision has become very important to the product selection and classification process. The problem becomes more challenging when it is necessary to detect defects or diseases in the product images. In literature, it was observed that when the fruit or vegetable image is treated as only one problem, efficiency is lower than when dividing it into sub-problems considering regions with similar appearance. Thus, in this paper, the target is to automate the detection and identification of visual defects in Brazil nuts by dividing the problem into two sub-problems (pulp and epidermis defects recognition) and by using color, shape and texture descriptors. First, the original image is segmented into two regions (one dark and one light). Then, First Order Descriptor, is applied to detect the presence or absence of defects in each region through the texture descriptor. Next, color, size and texture descriptors are used to the identification of each defect. This approach improves results obtained in previous research (Álvarez-Valera et al. [1]). We obtained an efficiency rate of 98.03 % with a processing time of 75 ms at worst and 51 at the best for every 3 images processed, unlike the previous attempt that had an efficiency rate of 91.79 % with a processing time of 130 ms. Finally, this approach can be applied in different types of products with other characteristics, since its inherent characteristics allows us to divide the original problem in two or more sub-problems.