Improving Defect Inspection Quality of Deep-Learning Network in Dense Beans by Using Hough Circle Transform for Coffee Industry

In this paper, we propose a novel Hough circle-assisting deep-network inspection scheme (HCADIS), aiming at identifying defects in dense coffee beans. The proposed HCADIS plays a critical role in a camera-based defect removal system to collect defective bean positions for picking all defects off. The idea of the HCADIS is to mix intermediate data from a deep network and a feature engineering method call Hough circle transform for utilizing advantages of both methods in inspecting beans. The Hough circle transform is adopted because it performs quite stable and bean shapes are highly close to circles in nature. A set of core mechanisms are designed for collaboration between the deep network and the Hough circle transform for precisely and accurately inspecting defective beans. Finally, we implement a prototype of the HCADIS and conduct experiments for testing the proposed scheme. The test results reveal that the HCADIS indeed successfully inspect defects among dense beans with superior performance in various metrics. This work provides industrial participants useful experiences for creating deep-learning solutions to bean products in coffee industries.