Title: Apple surface defect identification based on image analysis
Authors: Qunpo Liu; Yuxi Zhao; Jianjun Zhang; Ruxin Gao
Addresses: School of Electrical Engineering and Automation, Henan Polytechnic University, Jiaozuo, Henan, China; Henan International Joint Laboratory of Direct Drive and Control of Intelligent Equipment, Jiaozuo, Henan, China ' School of Electrical Engineering and Automation, Henan Polytechnic University, Jiaozuo, Henan, China ' School of Electrical Engineering and Automation, Henan Polytechnic University, Jiaozuo, Henan, China; Henan International Joint Laboratory of Direct Drive and Control of Intelligent Equipment, Jiaozuo, Henan, China ' School of Electrical Engineering and Automation, Henan Polytechnic University, Jiaozuo, Henan, China; Henan International Joint Laboratory of Direct Drive and Control of Intelligent Equipment, Jiaozuo, Henan, China
Abstract: The apple fruit defect detection is a necessary step before apples enter the market. When using deep learning to detect apple defects, apple defects are prone to miss detection and inaccurate positioning due to multiple convolutions and down-sampling. Therefore, this paper proposes YOLO-APPLE model. Three residual blocks in YOLOV3 were replaced with three dense blocks, and feature transfer between dense connected blocks was strengthened by combining average pooling to improve feature reuse, so as to reduce the rate of missed detection. Complete-IOU is used as the regression loss to locate the prediction frame more accurately. Secondly, K-means clustering algorithm was used for clustering apple defect dataset to obtain anchor boxes more consistent with apple defect and raise the efficiency of precision of the model. The results showed that the average precision of YOLO-APPLE model is 93.53%, and the detection speed is 43FPS, which can detect in real time.
Keywords: apple defect; YOLO-APPLE model; dense block; complete-IOU; k-means clustering algorithm.
DOI: 10.1504/IJCCPS.2022.124879
International Journal of Cybernetics and Cyber-Physical Systems, 2022 Vol.1 No.2, pp.169 - 183
Received: 14 Feb 2022
Accepted: 15 Feb 2022
Published online: 13 Aug 2022 *