AUSNUTZUNG VON FARBINFORMATIONEN ZUR SORTIERUNG VON NATURPRODUKTEN Peter Brückner 1, Thomas Lemanzyk 1, Günter Weber 2 1 Ilmenau University of Technology 2 MRB Automation GmbH, Ilmenau 29.- 30. September 2016 This work is encouraged by funds of BMWi under grant of ZIM AiF Project GmbH,, Germany 2013 Seite 1
Novel Approach on Optical Analysis and Multichannel Sorting of Grain Samples 1. Scope of Research 2. Image Acquisition 3. Image Processing Steps 4. Sorting Procedure 5. Mechanical Structure 6. Conclusions Fig. 1. Typical ingredients of a grain sample Seite 2
1. Scope of Research Identifying ingredients of a grain shipment is an important issue for: Avoidance of receiving toxic grain Define the necessity of cleaning steps Agreement of salary An automation physical sortation of grain fractions must fulfil three steps: membrane The first step is the recognition of all objects. The second step is the physical separation of each object classes. The third step is the weighting of the separated classes. (The rules and regulation of EU demands Mass percentage for each class) This Steps are done up to now by personal (subjective, slow). Seite 3
transmitted light lens camera 2. Image Acquisition The image acquisition is carried out by the color line scan camera and the conveyor belt with continuous velocity (v). hopper Advantage: All objects have a stabile uniform motion. v conveyer belt grain sample Disadvantage: We can only see the upper side of the kernels. diffusor Seite 4
2. Image Acquisition Principles of Color Line Scan Cameras (choice) beam splitter prism line array - blue red green blue lens line array - red line array - green Figure 3: Triple line solution (trilinear) One Chip with three times n-pixels (red, green, blue) Figure 2: Three sensor solution Beam splitter and three Chips (red, green, blue), each with n-pixels red + blue green Figure 4: Double line solution (bilinear) One Chip and two times n-pixels (red-blue and green) Seite 5
2. Image Acquisition Calculation of Line Rate for Square Pixels B pix = B o n pix B pix B o n pix - width of pixels in the object field - width of the object field - number of pixels L Pix = B Pix L Pix - length of pixels in the object field Z R = v L pix Z R v L Pix - line image rate - belt velocity - length of pixels in the object field Seite 6
2. Image Acquisition Numerical Example for Calculation of the Line Image Rate B pix = B o n pix B o = 153.6 mm n pix = 2048 L Pix = B Pix B pix = 0.075 mm L Pix = 0.075 mm square pixels Figure 5: Image of one kernel Z R = v L pix v = 750 mm/s Z R = 10,000 L/s Variable speed settings we obtain by synchronization of the camera speed (slave) and the belt velocity (master) with impulses of an incremental rotatory encoder. Seite 7
3. Image Processing Steps Color line scan camera Image acquisition Trilinear 4096 Pixel Color segmentation Object segmentation by Halcon Dot Net 227-dimensional feature vector Feature extraction by Halcon Dot Net Support Vector Machine Classification by Halcon Dot Net Seite 8
3. Image Processing Steps Segmentation- Separation of all objects from background Blue colored Belt: To achieve quite elementary segmentation of objects, a blue colored belt was chosen. Figure 6: Blue colored belt White illumination: The 3 LED- illumination sources, 2 incident and 1 transmitted light, has to deliver a flat and complete visible spectrum. Separation objects/background: p x, y = 1, b < (r, g) 0, otherwise (6) p x, y r, g, b Figure 7: Spectrum High Power LED- white -pixel in x-th column and y-th row -value of channels red, green and blue Seite 9
3. Image Processing Steps Segmentation- Separation of all objects from background Figure 8: Step-by-step separation of the objects from background Separation of color channels: RGB -> red, green, blue Median-Filter for blue channel (optional) Separation of: Regions: gray value of Pixels in red/green-channel > gray value of pixels in blue channel Background: (reverse) Closing (dilation + erosion) for segmented regions (optional) Joining for dissected objects in two following frames (MergeRegionLineScan) Joining of regions Seite 10
3. Image Processing Steps Feature extraction and Classification Selection of regions with reference to Size and Area (SelectShape: Size, Area) For each single region it s now necessary to calculate: Center of gravity, Image size (RegionFeatures: CenterPoint, ImageSize) Extraction of the feature vector (FeatureExtraction) Classifying of the objects (ClassifiyClassSvm) Each object is assigned to object class with maximum likelihood of correspondence (SVM classification). Storage of data Transmission of objects and images to the graphical user interface An Important Problem is: Natural products have a major inner class variability by minor distinctness of classes! Seite 11
3. Image Processing Steps Recognition rate r total = number of all objects correctly classified total number of objects 100% Ratio between the number of all objects correctly classified and the total number of objects. We reached: 94% total recognition rate 82% to 97% individual recognition rates Seite 12
4. Sorting Procedure First Stage: Separation of the main fraction (flawless kernels > 90 %). Channels (# 1 5) Refeeding Second Stage Separation of all other fractions in the first channel. Channel (# 1). Seite 13
4. Sorting Procedure Figure 9: Workflow of doublestage sorting principle. Seite 14
5. Mechanical structure 7 1 2 3 1 Hopper 2 Vibration conveyor 3 Conveyer belt 4 Ejector unit (major) 5 Container 6 Ejector unit (minor) 7 Refeeding 4 5 6 Figure 10: Mechanical structure of double-stage sorter Seite 15
5. Mechanical structure Figure 11: Prototype of multichannel sorting machine grainlab Seite 16
6. Conclusions Fast and multichannel sorting of natural objects is possible The new approach is an double stage sorting principle based on the existing of a main fraction. That is typical for grain and also typical for other food. No-one wants to produce food with less than 90 % good production. Build prototype with five channels in the 1 st stage, refeeding and one channel in the 2 nd stage- (outside position) Separation of the main fraction in all channels (5). Refeeding from all channels. Separation of all other fractions in the outside channel (1) Typical mass flow 50 grams per minute. Seite 17
Thank You for Your Attention! Seite 18