马铃薯畸形严重影响其商品价值,畸形剔除成为马铃薯产后售前的核心工作步骤之一.目前,该环节主要依靠人工目测进行,劳动消耗量大、投入成本高,精准、高效的自动检测技术亟待开发.近年来,机器视觉在物体外观及特性识别领域引起广泛关注,而马铃薯畸形属于马铃薯形态特征,故在获取马铃薯外形照片的基础上,采用改进的YOLOv3 算法对马铃薯畸形进行识别.使用注意力特征金字塔替换YOLOv3 算法中的特征金字塔,克服了特征融合过程中的干扰,增强网络的深层特征提取,并优化了特征表达,进而达到提升畸形检测精度、可靠性的目的.实验结果表明,改进的YOLOv3 算法相比改进前精确率提升 2.68%,F1 精度提升 2.31%,mAP提升 3.34%,针对深层特征的检测能力明显增强.该算法高效、精准,为马铃薯畸形检测提供了一种更优的智能检测方法.
Potato malformation seriously affects its commodity value,and the elimination of potato malformation has become one of the core work steps in the post production and pre-sale process of potatoes.At present,this process mainly relies on manual visual inspection,with large labor consumption and high investment cost.Therefore,accurate and efficient automatic detection technology needs to be developed urgently.In recent years,since machine vision has attracted extensive attention in the field of object appearance and feature recognition,and potato malformation belongs to potato morphological features,improved YOLOv3 algorithm was used to recognize potato malformation on the basis of obtaining potato appearance photos.By replacing the feature pyramid in the YOLOv3 algorithm with the feature pyramid of attention,the interference phenomenon in the process of feature fusion was overcome,the deep feature extraction of the network was enhanced,and the feature expression was optimized,so as to improve the accuracy and reliability of deformity detection.The experimental results showed that the accuracy of the improved YOLOv3 algorithm was improved by 2.68%compared with that before the improvement,F1 accuracy was 2.31%higher,mAP was 3.34%higher,and the detection ability for deep features was significantly enhanced.The algorithm is efficient and accurate,providing a better intelligent detection method for potato malformation detection.