Machine Vision-based Recognition Of Apple Stem/Calyx And Defects

Apples,as a popular fruit in the world wide,have many advantages,such as: strong ecological adaptation ability,high nutritional value and endurable storage.It is widely planted in C hina as a kind of industrial crops,with the total production ranking in the front of the world.However,due to the poor-processing techniques,the rate of the rotten apples is higher than 25%,severely decreasing the economic value.Hence,it is essential to improve the apple automatic recognition and defects detection techniques.Currently,machine vision-based techniques have been successfully app lied in recognizing the apples,for example: size,shape,color,internal quality and so on.But,because of shape similarity between the stem/calyx and defects,the detection of apple defects is still chal enging,which easily causes false detection.Concentrating on the machine vision-based apple defects detection problem,the aim of this paper is to recognize the apple stem/calyx and defects,with the main steps including object region segmentation,features extracting and defects recognizing.The main work is summarized below:(1)Building machine vision-based detection system.Image acquisition is the first key step in the image processing,which could guarantee the quality of captured image.So,an image processing system is proposed in this paper,including camera,light etc.(2)Apple image pre-processing methods.Due the camera shake and low-illu mination,the image contains severe noise,which effects the object region segme ntation and other post-processing steps.So,It is necessary to pre-processing the apple image before detection.Common ways to remove the noise is the median filtering and average filtering techniques.Experiments demonstrate that the result of median filtering method is better.(3)Object region segmentation.For a single view color apple image,the object regions can be segmented by using the methods including O tsu threshold segmentation method,canny edge detection method,hole filing method and edge exclusion method.Then a single color object region can be labeled with connected domain method.The segmentation results are proved to ideal.(4)Object region feature extraction.For a single object region,it is hard to recognize whether it belongs to stem,calyx or defects.Studies show that the stem,calyx and defects have some differences in shape,texture and color.So,by extracting features in those aspects,we can distinguish the stem,calyx and defects.(5)C lassier choosing.The choice of classier has a great effect on the recognition results.In this paper,we choose the most common used classier,including: Nearest Neighbor classifier,BP neural network,Ada Boost classier,SVM.Experiments show the recognition accuracy of these four classifier are 85.87% ? 89.13% ? 93.48% ?95.74%,respectively,and the result of SVM is the best.(6)On-line defect recognition system designing.In this paper,we incorporate the SVM with design tree as the classier in the on-line defect recognition system.The results figures out that the defection time of a single apple is 346 ms,which satisfies the real on-line need.