Rearch And Implementation Of Apple Detection System In Natural Scene

Fruit Harvesting Robot can effectively solve problems such as high labor intensity, labor shortages, low productivity and economic efficiency in agricultural picking work. The first step of the picking work is to fast identify ripe fruit and position them accurately. Recently, many researches on Harvesting Robot have stayed in theory or the experimental stage. This is mainly caused by low fruit recognition rate and slow processing speed which can hardly meet the requirement of real-time. Aiming at the above problems in this paper, ripe apples in natural scene are taken as study objects. Studying from the actual situation of the apple harvesting, automatic detection of ripe apples is realized in this paper. It can provide technical support for the design and development of apple Harvesting Robot in the future.The main work accomplished includes:(1) Due to the complex and changeable characteristics in natural scene and in order to get the overall recognition of the problem, a large number of apple pictures are taken to ensure the apple detection algorithm adaptive.(2) Based on analysis and research on various widely used image processing and feature extraction methods, a new detection approach suitable for apples in natural scene was proposed in this paper according to research object characteristics. Firstly, a vector median filter with 3×3 template was adopted to smooth the original image; Then using the image of R-G/2-B/2 in RGB color space, Otsu adaptive threshold method was applied to segment the images of ripe apples and background. Morphologic operation was employed to remove the random noise. After that the connecting areas were labeled through four connected sequence method. The remnant backgrounds were eliminated according to the areas of the labeled regions. Finally, Sobel operator is used to extract apple contour and an improved Hough transform is put in use to extract fruit feature parameters, which was able to solve the problems of low accuracy and slow speed in feature extracting.(3) An experiment was carried out to analyze the proposed apple detecting method by using pictures taken in natural scene. The experimental results showed that the overall recall and precision of the proposed approach is up to 93.5% and 95.9%. The recall reached the best value when individual apples were detected and the recall is 98.4%. Some bagged apples and yellow apples can also be detected by our method. Besides, apple images with background containing soil and the sky can also be well identified.(4) An apple detection system in Natural Scene was designed and implemented combined with C# and Matlab mixing programming technology, which promotes the development of vision technology for fruit harvesting robot.