农业工程学报
農業工程學報
농업공정학보
2013年
6期
157-165
,共9页
陈科尹%邹湘军%熊俊涛%彭红星%郭艾侠%陈丽娟
陳科尹%鄒湘軍%熊俊濤%彭紅星%郭艾俠%陳麗娟
진과윤%추상군%웅준도%팽홍성%곽애협%진려연
图像处理%模糊聚类%模拟退火%多尺度视觉显著性%粒子群算法%采摘机器人
圖像處理%模糊聚類%模擬退火%多呎度視覺顯著性%粒子群算法%採摘機器人
도상처리%모호취류%모의퇴화%다척도시각현저성%입자군산법%채적궤기인
image processing%fuzzy clustering%simulated annealing%multi-scale visual saliency%particle swarm%picking robot
准确分割水果图像是采摘机器人实现视觉定位的关键技术.该文针对传统模糊聚类对初始聚类中心敏感、计算量大和易出现图像过分割等问题,结合机器人的视觉特性,提出了一种基于多尺度视觉显著性改进的水果图像模糊聚类分割算法.首先,选择适当的颜色模型把彩色水果图像转换为灰度图像;然后对灰度图像做不同尺度的高斯滤波处理,基于视觉显著性的特点,融合了多个不同尺度的高斯滤波图像,形成图像聚类空间;最后,用直方图和模拟退火粒子群算法对图像的传统模糊聚类分割算法进行了改进,用改进的算法分别对采集到的100张成熟荔枝和柑橘图像,各随机选取50张,进行图像分割试验.试验结果表明:该方法对成熟荔枝和柑橘的图像平均果实分割率分别为95.56%和93.68%,平均运行时间分别为0.724和0.790 s,解决了水果图像过分割等问题,满足实际作业中采摘机器人对果实图像分割率和实时性的要求,为图像分割及其实时获取提供了一种新的基础算法,为视觉精确定位提供了有效的试验数据.
準確分割水果圖像是採摘機器人實現視覺定位的關鍵技術.該文針對傳統模糊聚類對初始聚類中心敏感、計算量大和易齣現圖像過分割等問題,結閤機器人的視覺特性,提齣瞭一種基于多呎度視覺顯著性改進的水果圖像模糊聚類分割算法.首先,選擇適噹的顏色模型把綵色水果圖像轉換為灰度圖像;然後對灰度圖像做不同呎度的高斯濾波處理,基于視覺顯著性的特點,融閤瞭多箇不同呎度的高斯濾波圖像,形成圖像聚類空間;最後,用直方圖和模擬退火粒子群算法對圖像的傳統模糊聚類分割算法進行瞭改進,用改進的算法分彆對採集到的100張成熟荔枝和柑橘圖像,各隨機選取50張,進行圖像分割試驗.試驗結果錶明:該方法對成熟荔枝和柑橘的圖像平均果實分割率分彆為95.56%和93.68%,平均運行時間分彆為0.724和0.790 s,解決瞭水果圖像過分割等問題,滿足實際作業中採摘機器人對果實圖像分割率和實時性的要求,為圖像分割及其實時穫取提供瞭一種新的基礎算法,為視覺精確定位提供瞭有效的試驗數據.
준학분할수과도상시채적궤기인실현시각정위적관건기술.해문침대전통모호취류대초시취류중심민감、계산량대화역출현도상과분할등문제,결합궤기인적시각특성,제출료일충기우다척도시각현저성개진적수과도상모호취류분할산법.수선,선택괄당적안색모형파채색수과도상전환위회도도상;연후대회도도상주불동척도적고사려파처리,기우시각현저성적특점,융합료다개불동척도적고사려파도상,형성도상취류공간;최후,용직방도화모의퇴화입자군산법대도상적전통모호취류분할산법진행료개진,용개진적산법분별대채집도적100장성숙려지화감귤도상,각수궤선취50장,진행도상분할시험.시험결과표명:해방법대성숙려지화감귤적도상평균과실분할솔분별위95.56%화93.68%,평균운행시간분별위0.724화0.790 s,해결료수과도상과분할등문제,만족실제작업중채적궤기인대과실도상분할솔화실시성적요구,위도상분할급기실시획취제공료일충신적기출산법,위시각정학정위제공료유효적시험수거.
The vision location system of the picking robot, which is an important part of the robot, is mainly used to detect the spatial position of the fruit and provide the motion control system of the robot with position information. Extracting the fruit waited for picking in a complex background by selecting an appropriate image segmentation technology provides us with the full assurance to obtain the position information of the fruit. So, aiming at the problems that the traditional fuzzy clustering is sensitive to the initial clustering centers and has large amounts of calculation and image over-segmentation, combining with the picking robot visual characteristics, an improved fuzzy clustering segmentation algorithm based on the multi-scale visual saliency for fruit image was put forward in this paper. First, a color model of the litchi and citrus image was discussed respectively, their diagrams of the R-I color model was expatiated, the fruit color image was converted into gray image by selecting a R-I color model; the gray image was processed with different scale Gaussian filters and the image clustering segmentation space was formed by blending all the different scale Gaussian filtering images according to the visual saliency, effect chart of the multi-scale visual saliency image algorithm was given based on R-I, and the over-segmentation problem most of the fruit image fuzzy clustering segmentation algorithms was solved. Second, the high dimensional clustering segmentation space based on pixels was changed into the low dimensional clustering segmentation space based on the histogram and the gray level by using the histogram method and the specific steps of image segmentation algorithm was given; the calculation of the fuzzy clustering image segmentation algorithm was greatly decreased and the fuzzy clustering image segmentation speed was improved. Furthermore, in the light of the problems that the fuzzy clustering algorithm easily fell into the local extreme value, the clustering center was optimized with the particle swarm algorithm based on simulated annealing, and the image clustering segmentation performance was improved. At the same time, the cooling strategy and state acceptance probability function of the particle swarm algorithm based on simulated annealing was nonlinearly reformed. Finally, the fuzzy clustering image segmentation algorithm based on multi-scale visual saliency of this paper was tested with 50 randomly selected images each of the 100 ripe litchi images and 100 ripe citrus images, and the contrast effect charts of the traditional and improved fruit image segmentation algorithms were given. The experimental results showed that:for the ripe litchi and citrus image, the average fruit segmentation rate of this method was 95.56% and 93.68%, and the average running time was 0.724 s and 0.790 s. The algorithm could meet the requirement of fruit image segmentation and real-time operation of the picking robot in the real picking activities;It has also provided a new basis algorithm for the image segmentation and its real-time research, and offered testing data for the vision accurate location of the picking robot.