Research On Tomato Detection Method Based On Computer Vision

Tomatoes are rich in nutrition which are one of the most commonly grown vegetables in Chinese agricultural production.But hand-picking tomatoes is a labor-intensive and time-consuming task.With the development of science and technology,much of work in agricultural production can be replaced by picking robots.First,the picking robot performs fruit detection using a computer vision system.Secondly,the mechanical arm is guided to perform the picking operation according to the detection results.In this process,the detection of fruit is the primary task and technical difficulty for picking robots,and the precision of detection is directly related to the robotic picking accuracy and efficiency.Due to the problem of inaccurate fruit detection by the vision system of tomato picking robots,this thesis investigates the method of tomato detection based on computer vision,with the following main studies:(1)A tomato detection method based on morphological reconstruction and threshold segmentation is investigated.Firstly,the features of different color spaces are analyzed,and the grayscale images of tomatoes are obtained by R-G chromatic aberration method.The transformed chromatic aberration image significantly enhances the contrast between the tomatoes and the background.Secondly,the images are subjected to median filtering and hole filling operations,thus removing noise from the image and eliminating tiny holes.Next,morphological reconstruction generates images with uniform gray scale characteristics and easy segmentation.Finally,the segmented images are adopted to extract the tomato edge using Canny edge detector,and the circular contours are fitted using the Hough transform,thus effectively detecting the ripe tomatoes.(2)A Mask R-CNN model based on Swin Transformer for tomato detection is proposed.Swin Transformer is used as the backbone network for better feature extraction.Multi-scale training techniques are shown to yield significant performance gains.The model not only effectively identifies tomato cultivars(normal-size tomatoes and cherry tomatoes)and tomato maturity stages(fully-ripened,half-ripened,and green),but also accurately detects and segments the tomatoes from the background environment.Through experimental analysis,the mean average accuracy(mAP)values of the model for tomato detection and segmentation are 89.4%and 89.2%,respectively.Experiments show that the mask images segmented by this model can accurately fit the tomatoes.