Application of Visual Classification in Automatic Screening of Granular Agricultural Products

[[abstract]]本研究主要是針對顆粒型農產品，例如脫膜蒜仁等所研發的自動篩選系統，整個自動篩選系統由取像裝置、電腦硬體、軟體辨識系統、打擊設備與運動機構所構成。利用機器視覺辨識選別，可以取代大量的人力工作，並且減少人為的疲勞及失誤所產生的誤差，進而提高辨識的精準度。我們首先會先分析目標物在RGB直方圖上的表現，利用設置閾值來濾除不要的背景，再利用建立樣品圖的方式去做目標物的色差分析，確認色差分析可以將目標物做階層式的分類後，再針對色差分析的缺點，以HSL的範圍分析方式找出在色差分析中不容易區分的部分，再將兩個結果組合後，最後我們可以得到一個辨識度達98%良品的方法。系統軟體部分使用C#搭配MIL影像函式庫撰寫，運作在64位元模式以提高運作速度，並以LineScan攝影機的固定取像周期為觸發條件，達到取像、處理與輸出控制同步的狀況。在硬體部分，我們還設計了可利用程式制定電流控制的LED驅動器，與可控制時間差的同步FPGA控制板，分別利用微處理器以及硬體描述語言達成控制目標。本系統的特點：（1）由於目標物篩選是在空中取像，大幅減少了處理背景的複雜度與時間，（2）不用擔心因長時間運作所弄髒的輸送帶之清潔問題，減少了維護成本，（3）採用複合式的判斷方法將帶膜物件再次篩選出來提高了辨識率，（4）自製光源驅動器使建置成本大幅降低，讓系統的實用性與被接受度大幅提高。最後本設備實際運作在蒜頭工廠，得到與設計測試過程相對應的成果，良品篩選率為98.7%，不良品篩選率為97.3%，高過於人工篩選的95.6%與93.2%，驗證了整體機構的設計與方法運作的可行性與實用性。[[abstract]]Thisresearchfocusesonadesign,theautomaticscreeningsystemforgranularagriculturalproducts,suchaspeeledgarlic.Thescreeningsystemiscomposedofanimagecapturedevice,computersoftwareandhardwareidentificationsystem,aclassificationdeviceandamovementdevice.Theadvantagesofmachinevisualidentificationareasfollows.First,lotsofmanpowercanbereplacedwithmachines.Then,theerrorscausedbyman-madenegligencecanbehighlyreduced.Consequently,theaccuracyofidentificationcanbeimproved.Inthebeginning,weanalyzedtheperformanceonRGBhistogram.Unnecessarybackgroundwasremovedviasettingthresholdvalue.Then,weanalyzedcolordifferenceofthetargetbysettingupthesampleimages.Afterthat,thelevelsofthetargetwereclassified.However,therearestilldisadvantagesincolordifferenceanalysis.Therefore,HSL（Hue,Saturation,Lightness）isusedtohelpconfirmthesectionswhicharenoteasilydistinguished.Combiningthetwomethods,an98%accuracyofscreeningqualifiedproductswasachieved.ThesoftwareiscompletedwithC#andMIL,operatingina64-bitmodetoimprovethespeed.ThefixedcycleofimagecaptureinLineScancameraisusedasatriggertoachievethesynchronizationofimagecapture,processingandoutputcontrol.Asforhardware,wedesignedaprogrammableconstantcurrentLEDdriverandaprogrammabletimesynchronizationFPGAcontrolboard.ByusingMCUandVHDL,wecansuccessfullycontrolthetarget.Thefeaturesofthesystemareasfollows:1.Thecomplexityandtime-consumingaresignificantlyreducedinremovingbackgroundduetothereasonthattheimagecaptureisdoneintheair.2.Themaintenancecostismuchlowerbecauseweareabletosavelotsofmoneyincleaningconveyors.3.Theaccuracyishighlyincreasedbyusingthecomplexjudgementmethods.4.ThepracticalityandacceptancearehighlyincreasedbecausethecostismuchlowerwiththeLEDdrivermadeourselves.Finally,thesystemwasactuallyoperatedinagarlicfactory.Itgotaresultwhichmettheexperimentsinourdesign—98.7%accuracyofscreeningqualifiedproductsand97.3%ofscreeningdefectiveproducts.Theaccuracyratesresultedfromourdesignareobviouslysuperiorto95.6%and93.2%frommanualscreening.Theresultfullyverifiedthefeasibilityandpracticabilityofourdesign