CN109030504A - The panorama potato defect detecting device and method combined based on actual situation imaging - Google Patents

Abstract

The invention discloses a panoramic potato defect detection device and method based on the combination of virtual and real imaging, including a conveying device, a darkroom detection device, a control module and an intelligent algorithm defect identification module, the control module and the conveying device, the darkroom detection device and an intelligent algorithm defect The identification modules are connected, and the control module controls the start and stop of the conveying device and the darkroom detection device, and at the same time controls the conveying device according to the results of the intelligent algorithm defect identification module to complete the classification. The darkroom detection device includes a camera, a ring LED light, a push rod mechanism, an upper box, a lower box, a transparent plate, a rack and pinion mechanism, a spring, a pole, a support cover, a reflector mechanism, and a bottom LED light. The invention patent uses a single camera, which can simultaneously image a real image on the top of the potato and virtual images of four mirrors, and complete the panoramic defect detection without turning the potato.

Description

The panorama potato defect detecting device and method combined based on actual situation imaging

Technical field

The invention belongs to agricultural product quality detection technique fields, and in particular to a kind of that the panorama combined is imaged based on actual situation Potato defect detecting device and method, according to the shape and external appearance characteristic of potato, in the case where not stirring potato, The panoramic vision detection that actual situation imaging combines is carried out to potato, to complete the real-time defects detection and classification of potato.

Background technique

China is maximum potato producing country in the world, and potato is classified as the fourth-largest staple food grain.But for Ma Ling The detection of potato defect, most of manually sense organ that also rests on carry out the identification judgement stage.Potato is completed by manual type Classification occur amount of labour used famine with the harvest season first is that wasting a large amount of labours, second is that subjective assessment is by personal warp It tests, testing result consistency is poor, and low efficiency, it is difficult to which the requirement for meeting high standard classification is unfavorable for realizing automation.With The fast development of artificial intelligence depth learning technology, the defect of potato can be recognized automatically by carrying out detection using machine vision.

At present the grading plant of potato used in major part potato processing enterprise generally all simply by weight into Row classification obtains potato quality data to be measured using weight sensor or pressure sensor, then by analog circuit or Person's digital circuit is classified.But these devices are for having potato of surface defect, such as germination, shagreen, breakage, deformity etc. Defect can not be handled.Thus, it is also necessary to configuring manpower in advance chooses substandard products, is classified further according to potato quality, It cannot achieve complete mechanization or automation in this way.

Although the potato stage division and equipment that are now based on computer vision are increasingly becoming hot spot, generally it is only limited to Laboratory research takes pictures to it using single camera, and under the premise of not stirring potato, can not obtain horse to be measured The panoramic surface information of bell potato, there are the probability of missing inspection；Or have and taken pictures using multiple video cameras to potato to be checked, but It is to increase the cost and its complexity of detection device, while increasing detection time.

In order to solve the defect of existing potato grading plant, is not increasing number of cameras and its stirring potato to be checked Under the premise of, propose a kind of panorama potato defect inspection method and its device for being imaged and combining based on actual situation.

There are some Patents documents to report by domestic Searches of Patent Literature discovery, mainly have following some:

1, notification number is 104056790 A of CN, and entitled " a kind of potato intelligent separation method and device " utility model is special Benefit, solving existing potato classifying equipoment can be according to weight grading, but uses the technical solution of single camera detection, cannot Realize all standing on potato surface to be checked.

2, notification number is 104941922 B of CN, a kind of entitled " potato grading control based on machine vision technique Device " patent of invention may be implemented the quick detection sorting of impurity, different quality potatos, be rejected using air ejector miscellaneous Matter controls potato and is oriented to the collision angle between driving lever to reduce impact force, uses one according to the detection transverse diameter of potato A or multiple guiding mechanisms reject potato, realize effective sorting of potato to be fractionated, but the sorter uses The throwing mode of getting rid of is classified, and there are the risks of cuticular breakdown, in addition, there are potatos to be checked for used video camera shooting style Surface cannot be completely covered, and there is blind area of taking pictures.

3, notification number is 203732461 A of CN, a kind of entitled " horizontal feed for Potato Quality Image Acquisition At the uniform velocity turnover device " utility model patent designs a kind of horizontal feed for Potato Quality Image Acquisition and at the uniform velocity turns over Potato external sort non-destructive testing may be implemented in rotary device, but the device is needed by concave cylindrical material supporting roller mechanism Potato is overturn, the breakage of potato epidermis is be easy to cause, and for various sizes of potato, rollover effect is deposited In difference, a possibility that leading to missing inspection.

4, notification number be 106238342 A of CN, it is entitled " panoramic vision potato sorting and defect detecting device and its Sort detection method " patent of invention, the surface information of potato to be checked is obtained from 6 different directions using 6 video cameras, although Panoramic information can be obtained, used camera was excessive at that time, and installation cost is high；In addition, being imaged in order to prevent during taking pictures The light filling of machine is interfered, and 6 video cameras is needed successively to shoot, detection time length, low efficiency.

Although above patent document proposes the method for sorting and potato grading plant of potato, although some equipment energy It is classified according to external appearance characteristic, but since only with a camera, detection is not comprehensive, and there are the probability of missing inspection, or needs to turn over Turn potato, is easily damaged epidermis；Although what is had can be carried out panorama detection, use video camera excessive, it is at high cost, and by It is limited in testing principle, detection efficiency is low.

Summary of the invention

The present invention in order to solve shortcoming in the prior art, provide a kind of detection comprehensively, prevent missing inspection situation substantially Generation and do not damage potato epidermis based on the actual situation panorama potato defect detecting device that combines of imaging and method.

In order to solve the above technical problems, the present invention adopts the following technical scheme: the panorama horse combined is imaged based on actual situation Bell potato defect detecting device, including the imaging of conveying device, actual situation combine formula darkroom detection device and control cabinet；

Conveying device include feeding conveying mechanism, conveying body, high pressure draught nozzle mechanism, defect potato lower skateboard and Defect potato collecting box；Output end on rear side of feeding conveying mechanism combines on front side of the detection device of formula darkroom with actual situation imaging Input port connection, actual situation imaging combine the input terminal on front side of delivery outlet and conveying body on rear side of the detection device of formula darkroom Connection；The left side of conveying body is arranged in high pressure draught nozzle mechanism, and defect potato lower skateboard is located at outfeed belt conveyor The right side of structure is simultaneously oppositely arranged with high pressure draught nozzle mechanism or so, and defect potato lower skateboard is obliquely installed in right low left high, The lower section of defect potato lower skateboard right edge is arranged in defect potato collecting box；The rear side output end of conveying body is Qualified potato outlet；

Be provided with control module and intelligent algorithm defect recognition module in control cabinet, control module by Industrial Ethernet respectively with Conveying device combines formula darkroom detection device with actual situation imaging and connects, and intelligent algorithm defect recognition module and control module connect It connects.

It includes vertically disposed four columns that actual situation imaging, which combines formula darkroom detection device, is provided between four columns The lower box of open top, the upper end of each column have been bolted a connector, have been arranged between four connectors There is the upper box of bottom opening, upper box and lower box are oppositely arranged up and down and have gap, and actual situation imaging combines formula darkroom The delivery outlet of input port and rear side on front side of detection device is separately positioned on the lower front and back lower of upper box, input port It is arranged with delivery outlet correspondence；

There are four the branch cover for seat of open top, four branch cover for seat are respectively adjacent to four sides of lower box for bottom setting in lower box Wall, the line of four branch cover for seat are diamond shape, and a spring is provided in each cover for seat, and spring upper end is higher than branch cover for seat, often It is inserted with the strut being arranged in vertical in a spring, the gear with spring upper end press cooperation is provided on strut Ring, strut lower end are extend into branch cover for seat and apart from branch cover for seat bottom 2-5cm, the right side of the lower part of left side strut, right side strut The left side of lower part, be provided with rack structure on front side of the rear side of lower part of front side strut and the lower part of rear side strut, often A rack structure is engaged with a first gear, the central coaxial of first gear to being fixedly installed a bull stick, bull stick Both ends are rotatably connected on the opposite two sides inner wall of lower box by bearing, and the medium position of bull stick is arranged reflective by mounting rack The reflective mirror that face tilts upward；The upper end of four struts is horizontally disposed one piece of transparent panel, and transparent panel is in rectangle and and lower box There is the gap 1-3mm between four side walls；It is respectively arranged with a potato gear on the strut on left side and strut and right side and pushes away machine Structure, two potato gear pushing mechanism note constructions are identical and are symmetrical set.

Potato gear pushing mechanism positioned at right side includes support plate, and support plate is fixedly connected in lower box outer right wall, The lower surface of support plate is provided with gear and pushes away stepper motor, and the main shaft that gear pushes away stepper motor passes through support plate vertically upward and coaxially connects It is connected to second gear, support plate upper surface is rotatably connected to third gear, and second gear and third gear are respectively positioned on transparent panel The right side of right edge, third gear are engaged with second gear, and the center of third gear upper surface is fixedly connected with horizontally disposed Gear push rod, keep off push rod lower surface be higher than transparent panel upper surface 1-3mm, keep off center of the both ends of push rod about third gear Symmetrically；The gear push rod of two potato gear pushing mechanisms has 2-10mm when turning in straight line between two gear push rods Gap, gear push rod gap across upper box and lower box in rotation.

Video camera is installed, the periphery of video camera is provided with annular LED lamp at the center of upper box inner top；Lower box Interior bottom is provided with the lower LED light irradiated upward.

Feeding conveying mechanism and conveying body include that multiple groups are successive or one group of belt conveyor, belt Conveying mechanism includes bracket, and bracket left or right side is provided with decelerating motor, is rotatably connected to drive roll and driven voller on bracket, Drive roll and driven voller are horizontally disposed in left-right direction, are rotatablely connected between drive roll and driven voller by the belt of annular, Several are arranged at intervals in the middle part of belt outer surface for accommodating the groove of single potato, groove, input port and delivery outlet It is arranged in straight line, the output end of decelerating motor is provided with driving pulley, and one end of drive roll is provided with driven pulley, actively Belt wheel is connected with driven pulley by V V belt translation.

Control module includes embedded-type ARM single-chip microcontroller, I/O submodule, Industrial Ethernet communication submodule, display submodule Block, light warning submodule and computer；Computer be intelligent algorithm defect recognition module operation physical support, computer with Industrial Ethernet communication submodule is connected；Embedded-type ARM single-chip microcontroller communicates submodule with Industrial Ethernet, I/O submodule, shows Show that submodule is connected with light warning submodule, this is the key submodule of control module；I/O submodule respectively with deceleration Motor, gear push away step motor control signal and are connected, and corresponding gear is successively controlled according to control flow, and to push away stepper motor orderly Movement, the I/O submodule is also connected with the electric control valve in high pressure draught nozzle mechanism, according to recognition result, open or Person closes electric control valve；The I/O submodule is also connected with and controls the opening and closing of annular LED light and lower LED light；The computer is logical It crosses usb bus mode to be connected with video camera, by Industrial Ethernet communication mode, computer is by embedded-type ARM single-chip microcontroller Instruction, starts taking pictures for video camera, and emphasis of taking pictures for the first time obtains parts of images on potato, and emphasis of taking pictures for the second time obtains horse Parts of images under bell potato, and photo is deposited into computer.

Intelligent algorithm defect recognition module includes image preprocessing submodule, image segmentation submodule and multichannel convolutive net String bag module composition；Described image pretreatment submodule pre-processes the image taken pictures twice, promotes the quality of image, locates Image after having managed is sent to image segmentation submodule；Described image segmentation submodule handles image of taking pictures for the first time, The real image edge of potato is obtained, and is split, zooms to the fixed dimension of 256 × 256 pixels, described image divides submodule Block handles image of taking pictures for the second time, obtains four virtual image edges of potato, and is split, zooms to 128 × 128 One real image and four virtual images are transported to the multi-pass again by the fixed dimension of pixel, described image segmentation submodule simultaneously Road convolutional network submodule is identified, detects potato sprouting, shagreen, breakage or a series of defects of deformity, and tie checking Fruit is sent to control module by Industrial Ethernet.

Detection method based on the panorama potato defect detecting device that actual situation imaging combines, includes the following steps,

(1), into the preparation stage before the detection device of darkroom；The start and stop of decelerating motor, feeding conveying are controlled by control module Ensure that only one potato to be transported occurs in each groove on the belt of mechanism, prevents multiple potatos to be checked simultaneously Enter darkroom detection device；

(2), the darkroom detection device detent stage；Before entering darkroom detection device, the gear of potato gear pushing mechanism pushes away potato Under the instruction of control module, gear pushes away stepper motor driving gear push rod and arranges in front wide and rear narrow "eight" shape stepper motor, makes After potato enters on transparent panel in the input port by upper box, rapid detent to position to be detected；When potato opens Begin before detection, gear pushes away stepper motor under the instruction of control module, and gear pushes away stepper motor driving gear push rod and returns to reset condition, is in " two " font arrangement, takes pictures convenient for video camera；

After potato enters on transparent panel, due to the effect of gravity, transparent panel is moved downward, pressing spring under four struts, Rack structure driving first gear rotation on strut, the coaxial bar rotation with first gear, bar rotation drive reflective mirror Rotation, the amplitude that the big potato of quality moves down transparent panel is big, and the amplitude that the small potato of quality moves down transparent panel is small；From And reflective mirror is made to carry out angle adjustment according to the gravity size of potato, be conducive to video camera and capture four reflective mirror internal reflections Potato lower part four virtual images；

(3), the actual situation image checking stage；It first carries out taking pictures for the first time, obtains the image of part on potato to be detected, the image For real image, the embedded-type ARM single-chip microcontroller in control module issues the annular LED lamp that top in upper box is opened in order, under closing The lower LED light of the bottom of box, after the top real image for obtaining potato to be measured, using image segmentation submodule to the real image at Reason, extracts the real image edge of potato, and be split, zoom to the fixed dimension of 256 × 256 pixels；Then second is carried out Secondary to take pictures, the embedded-type ARM single-chip microcontroller in control module issues the lower LED light that lower box bottom is opened in order, closes upper box In top annular LED lamp, the image of potato lower part in reflective mirror is reflected in four reflective mirrors, which is the virtual image, is taken the photograph After camera obtains four virtual images of the lower part of potato to be measured, four virtual image edges of potato are extracted respectively, and be split, Zoom to the fixed dimension of 128 × 128 pixels；Image segmentation submodule again will a treated real image and four virtual images Be transported to the multichannel convolutive network submodular simultaneously and identified, detection potato whether have germination, shagreen, breakage or The defect of deformity, and control module is sent by Industrial Ethernet by inspection result；

(4), darkroom detection device and sorting stage are released；After the completion of detection, gear pushes away stepper motor under the instruction of control module, Gear pushes away stepper motor and rotates backward two gear push rods of driving into " V " font arrangement, potato is released backward defeated on upper box Outlet, potato are moved on conveying body；For flawless potato under the conveying of conveying body, directly It is moved to next process to be handled, for defective potato, control module instruction is opened in high pressure nozzle mechanism Electric control valve, high pressure nozzle mechanism sprays the side that potato is pushed to conveying body by high pressure draught, from defect potato Lower skateboard slides into defect potato collecting box.

Display sub-module in step (3) in control module is according to detection as a result, showing normal horse of current operation phase The number of bell potato, the number of all kinds of defect potatos, and detection sum；When device has it is abnormal when, pass through light warning submodule Block is warned；

Image preprocessing submodule, image segmentation submodule and multichannel volume in step (3) intelligent algorithm defect recognition module Product network submodular specific work process be；

Top potato real image after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel convolutive A1 convolutional layer in network submodular, after the operation of 5 × 5 window convolutions, 15 width images of 254 × 254 pixels of generation, then by The pond A2 layer in multichannel convolutive network submodular carries out compression processing, generates 15 width images of 127 × 127 pixels, later, Second of convolution operation is carried out, the A3 convolutional layer being sent into multichannel convolutive network submodular is operated using 3 × 3 window convolutions Afterwards, 60 width images of 125 × 125 pixels are generated, then compression is carried out by the pond the A4 layer in multichannel convolutive network submodular Reason generates 60 width images of 63 × 63 pixels, further, at the full articulamentum of A5 in multichannel convolutive network submodular Reason exports the vector of 4096 dimensions, further, handles by the full articulamentum of A6 in multichannel convolutive network submodular, defeated The vector of 1024 dimensions out；

Lower part the first potato virtual image after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel B1 convolutional layer in convolutional network submodule generates 15 width images of 126 × 126 pixels after the operation of 5 × 5 window convolutions, Compression processing is carried out by the pond the B2 layer in multichannel convolutive network submodular again, generates 15 width images of 63 × 63 pixels, it Afterwards, second of convolution operation is carried out, the B3 convolutional layer being sent into multichannel convolutive network submodular is grasped using 3 × 3 window convolutions After work, 60 width images of 61 × 61 pixels are generated, then compression is carried out by the pond the B4 layer in multichannel convolutive network submodular Reason generates 60 width images of 31 × 31 pixels, further, at the full articulamentum of B5 in multichannel convolutive network submodular Reason exports the vector of 1024 dimensions, further, handles by the full articulamentum of B6 in multichannel convolutive network submodular, defeated The vector of 256 dimensions out；

Lower part the second potato virtual image after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel C1 convolutional layer in convolutional network submodule generates 15 width images of 126 × 126 pixels after the operation of 5 × 5 window convolutions, Compression processing is carried out by the pond the C2 layer in multichannel convolutive network submodular again, generates 15 width images of 63 × 63 pixels, it Afterwards, second of convolution operation is carried out, the C3 convolutional layer being sent into multichannel convolutive network submodular is grasped using 3 × 3 window convolutions After work, 60 width images of 61 × 61 pixels are generated, then compression is carried out by the pond the C4 layer in multichannel convolutive network submodular Reason generates 60 width images of 31 × 31 pixels, further, at the full articulamentum of C5 in multichannel convolutive network submodular Reason exports the vector of 1024 dimensions, further, handles by the full articulamentum of C6 in multichannel convolutive network submodular, defeated The vector of 256 dimensions out；

The lower part third potato virtual image after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel D1 convolutional layer in convolutional network submodule generates 15 width images of 126 × 126 pixels after the operation of 5 × 5 window convolutions, Compression processing is carried out by the pond the D2 layer in multichannel convolutive network submodular again, generates 15 width images of 63 × 63 pixels, it Afterwards, second of convolution operation is carried out, the D3 convolutional layer being sent into multichannel convolutive network submodular is grasped using 3 × 3 window convolutions After work, 60 width images of 61 × 61 pixels are generated, then compression is carried out by the pond the D4 layer in multichannel convolutive network submodular Reason generates 60 width images of 31 × 31 pixels, further, at the full articulamentum of D5 in multichannel convolutive network submodular Reason exports the vector of 1024 dimensions, further, handles by the full articulamentum of D6 in multichannel convolutive network submodular, defeated The vector of 256 dimensions out；

The 4th potato virtual image of lower part after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel E1 convolutional layer in convolutional network submodule generates 15 width images of 126 × 126 pixels after the operation of 5 × 5 window convolutions, Compression processing is carried out by the pond the E2 layer in multichannel convolutive network submodular again, generates 15 width images of 63 × 63 pixels, it Afterwards, second of convolution operation is carried out, the E3 convolutional layer being sent into multichannel convolutive network submodular is grasped using 3 × 3 window convolutions After work, 60 width images of 61 × 61 pixels are generated, then compression is carried out by the pond the E4 layer in multichannel convolutive network submodular Reason generates 60 width images of 31 × 31 pixels, further, at the full articulamentum of E5 in multichannel convolutive network submodular Reason exports the vector of 1024 pixels, further, handles by the full articulamentum of E6 in multichannel convolutive network submodular, defeated The vector of 256 dimensions out；

Finally, exporting vector, the B6 of 1024 dimensions to the full articulamentum of A6 by the A7 tether layer in multichannel convolutive network submodular The full articulamentum of vector, D6 that the full articulamentum of vector, C6 that full articulamentum exports 256 dimensions exports 256 dimensions exports 256 dimensions The vector that the full articulamentum of vector sum E6 exports 256 dimensions executes set add operation, generates lacking for characterization potato panoramic surface Fall into characteristic information, be the vector of 2048 dimensions, then from soft the recurrences layer of the A8 in multichannel convolutive network submodular export 5 tie up to Amount indicates that potato to be detected belongs to the probability density distribution of normal, germination, shagreen, breakage or lopsided surface defect, to distinguish Potato to be checked is known with the presence or absence of defect；

Multichannel convolutive network submodular includes 5 channels, can disposably input 1 potato upper part real image image to be detected With part virtual images under 4 width；The off-line training sample database of multichannel convolutive network submodular can increase sample size；Cause And the germination of potato, shagreen, breakage, lopsided type can further be segmented with the increase of sample size；Multichannel volume Product network submodular can be both trained update in use by user, also can choose regular by device manufacturer It updates；The present apparatus supports the multichannel convolutive network submodular of multi version, can be carried out by end user according to practical application scene Autonomous selection.

By adopting the above technical scheme, the present invention has following technical effect that

It is of the invention panorama to be combined based on actual situation imaging potato defect is detected, by single camera, do not turning over Under the premise of turning potato to be checked, it can be realized and take pictures detection in panorama, one side is it is possible to prevente effectively from potato epidermis touches Wound, breakage, still further aspect are reduced installation cost, and improve detection efficiency, are rolled up using multichannel by single camera Product neural network, promotes the accuracy rate of detection.The present invention can be used for potato agricultural product defect on-line real-time measuremen, for being promoted The development of China's potato deep processing of farm products has great importance, and market application prospect is preferable.

Detailed description of the invention

Fig. 1 is overall structure diagram of the invention；

Fig. 2 is the upper box and lower box external structure schematic diagram of darkroom detection device in the present invention；

Fig. 3 is the bottom view of the upper box of darkroom detection device in the present invention；

Fig. 4 is the lower box schematic diagram of internal structure of darkroom detection device in the present invention；

Fig. 5 is the operating position schematic diagram for keeping off push rod；

Fig. 6 is that the schematic diagram of taking pictures combined is imaged to small size potato actual situation to be checked；

Fig. 7 is that the schematic diagram of taking pictures combined is imaged to large scale potato actual situation to be checked

Fig. 8 is reflective mirror (plane mirror and convex mirror) optics virtual image schematic diagram；

Fig. 9 is control module logical connection schematic diagram；

Figure 10 is multichannel convolutive network submodular schematic diagram.

Specific embodiment

It is of the invention that the panorama potato defect detecting device combined, packet are imaged based on actual situation as shown in Fig. 1-Fig. 8 Include conveying device, actual situation imaging combines formula darkroom detection device 37 and control cabinet 1；

Conveying device includes feeding conveying mechanism, conveying body, high pressure draught nozzle mechanism 2, defect potato lower skateboard 3 With defect potato collecting box 4；Output end on rear side of feeding conveying mechanism combines formula darkroom detection device 37 with actual situation imaging The input port 5 of front side connects, before actual situation imaging combines the delivery outlet 6 and conveying body of 37 rear side of formula darkroom detection device The input terminal of side connects；The left side of conveying body, defect potato lower skateboard 3 is arranged in high pressure draught nozzle mechanism 2 In conveying body right side and be oppositely arranged with high pressure draught nozzle mechanism 2 or so, defect potato lower skateboard 3 be in left height The lower section of 3 right edge of defect potato lower skateboard is arranged in right low dip setting, defect potato collecting box 4；Outfeed belt conveyor The rear side output end of structure is qualified potato outlet；2 air inlet of high pressure draught nozzle mechanism and plant source or air compressor machine phase Connection (for this field routine techniques, does not identify) in figure.

Control module and intelligent algorithm defect recognition module are provided in control cabinet 1, control module passes through Industrial Ethernet It combines formula darkroom detection device 37 with conveying device and actual situation imaging respectively to connect, intelligent algorithm defect recognition module and control Module connection.

It includes vertically disposed four columns 7 that actual situation imaging, which combines formula darkroom detection device 37, is set between four columns 7 It is equipped with the lower box 8 of open top, a connector 9, four connectors 9 have been bolted in the upper end of each column 7 Between be provided with the upper box 10 of bottom opening, upper box 10 and about 8 lower box are oppositely arranged and have gap, and actual situation is imaged The delivery outlet 6 of the input port 5 and rear side that combine 37 front side of formula darkroom detection device is separately positioned under the front side of upper box 10 Portion and back lower, input port 5 and the setting of 6 correspondence of delivery outlet；

There are four the branch cover for seat 11 of open top, four branch cover for seat 11 are respectively adjacent to the four of lower box 8 for bottom setting in lower box 8 A side wall, the line of four branch cover for seat 11 are diamond shape, and a spring 12 is provided in each cover for seat 11, and 12 upper end of spring is high In being inserted with the strut being arranged in vertical 13 in branch cover for seat 11, each spring 12, it is provided on strut 13 and bullet The baffle ring 14 of 12 upper end of spring press cooperation, 13 lower end of strut extend into branch cover for seat 11 and apart from 11 bottom 2-5cm of branch cover for seat, left The right side of the lower part of side strut 13, the left side of the lower part of right side strut 13, front side strut 13 lower part rear side and rear side branch It is provided with rack structure 15 on front side of the lower part of bar 13, each rack structure 15 is engaged with a first gear 16, and first The central coaxial of gear 16 is rotatably connected on lower box by bearing 18 to a bull stick 17, the both ends of bull stick 17 are fixedly installed On 8 opposite two sides inner walls, the reflective mirror 20 that reflective surface tilts upward is arranged by mounting rack 19 in the medium position of bull stick 17；Four The upper end of root strut 13 is horizontally disposed one piece of transparent panel 21, and transparent panel 21 has in rectangle and between 8 four side wall of lower box The gap 1-3mm；A potato gear pushing mechanism, two horse bells are respectively arranged on left side and strut 13 and the strut on right side 13 Potato gear pushing mechanism note construction is identical and is symmetrical set.

Potato gear pushing mechanism positioned at right side includes support plate 22, and it is outer that support plate 22 is fixedly connected on 8 right side of lower box On wall, the lower surface of support plate 22 is provided with gear and pushes away stepper motor 23, and the main shaft that gear pushes away stepper motor 23 passes through branch vertically upward Fagging 22 is simultaneously coaxially connected with second gear 24, and 22 upper surface of support plate is rotatably connected to third gear 25,24 He of second gear Third gear 25 is respectively positioned on the right side of the right edge of transparent panel 21, and third gear 25 is engaged with second gear 24, third gear The center of 25 upper surfaces is fixedly connected with horizontally disposed gear push rod 26, and the lower surface of gear push rod 26 is higher than 21 upper surface of transparent panel 1-3mm keeps off central symmetry of the both ends of push rod 26 about third gear 25；The gear push rod 26 of two potato gear pushing mechanisms exists When turning in straight line, the gap with 2-10mm between two gear push rods 26, gear push rod 26 passes through top box in rotation Gap between body 10 and lower box 8.

Video camera 27 is installed, the periphery of video camera 27 is provided with annular LED lamp at the center of 10 inner top of upper box 28；Bottom is provided with the lower LED light (not shown) irradiated upward in lower box 8.

Feeding conveying mechanism and conveying body include that multiple groups are successive or one group of belt conveyor, belt Conveying mechanism includes bracket 29, and 29 left or right side of bracket is provided with decelerating motor 30, is rotatably connected to drive roll on bracket 29 31 and driven voller 32, drive roll 31 and driven voller 32 be horizontally disposed in left-right direction, between drive roll 31 and driven voller 32 lead to The belt 33 for crossing annular is rotatablely connected, several are arranged at intervals in the middle part of 33 outer surface of belt for accommodating single potato Groove (is not illustrated) in figure, and groove, input port 5 and delivery outlet 6 are arranged in straight line, and the output end of decelerating motor 30 is provided with Driving pulley 34, one end of drive roll 31 are provided with driven pulley 35, driving pulley 34 and driven pulley 35 and are driven by V band 36 Connection.

As shown in Figure 9 and Figure 10, control module includes embedded-type ARM single-chip microcontroller, I/O submodule, Industrial Ethernet communication Submodule, display sub-module, light warning submodule and computer；Computer is the object of intelligent algorithm defect recognition module operation Carrier is managed, computer is connected with Industrial Ethernet communication submodule；Embedded-type ARM single-chip microcontroller communicates submodule with Industrial Ethernet Block, I/O submodule, display sub-module are connected with light warning submodule, this is the key submodule of control module；I/O Module respectively with decelerating motor 30, gear push away stepper motor 23 control signal be connected, successively controlled according to control flow corresponding to Gear push away stepper motor 23 and orderly act, the I/O submodule is also connected with the electric control valve in high pressure draught nozzle mechanism 2 It connects, according to recognition result, turns on or off electric control valve；The I/O submodule is also connected with and controls annular LED light 28 under The opening and closing of LED light；The computer is connected by usb bus mode with video camera 27, by Industrial Ethernet communication mode, Computer presses the instruction of embedded-type ARM single-chip microcontroller, and starting video camera 27 is taken pictures, and emphasis of taking pictures for the first time obtains potato top Partial image, emphasis of taking pictures for the second time obtains parts of images under potato, and photo is deposited into computer.

Intelligent algorithm defect recognition module includes image preprocessing submodule, image segmentation submodule and multichannel convolutive net String bag module composition；Described image pretreatment submodule pre-processes the image taken pictures twice, promotes the quality of image, locates Image after having managed is sent to image segmentation submodule；Described image segmentation submodule handles image of taking pictures for the first time, The real image edge of potato is obtained, and is split, zooms to the fixed dimension of 256 × 256 pixels, described image divides submodule Block handles image of taking pictures for the second time, obtains four virtual image edges of potato, and is split, zooms to 128 × 128 One real image and four virtual images are transported to the multi-pass again by the fixed dimension of pixel, described image segmentation submodule simultaneously Road convolutional network submodule is identified, detects potato sprouting, shagreen, breakage or a series of defects of deformity, and tie checking Fruit is sent to control module by Industrial Ethernet.Intelligent algorithm defect recognition module operates in computer environment, controls mould Block, intelligent algorithm defect recognition module and computer are this field common technique, and concrete model is not listed.

Detection method based on the panorama potato defect detecting device that actual situation imaging combines, comprising the following steps:

(1), into the preparation stage before the detection device of darkroom；The start and stop of decelerating motor 30 are controlled by control module, feeding is defeated Sending ensures that only one potato to be transported occurs in each groove on the belt 33 of mechanism, prevent multiple potatos to be checked Darkroom detection device is entered simultaneously；

(2), the darkroom detection device detent stage；Before entering darkroom detection device, the gear of potato gear pushing mechanism pushes away potato For stepper motor 23 under the instruction of control module, gear pushes away the driving gear push rod 26 of stepper motor 23 in front wide and rear narrow "eight" shape cloth It sets, so that after potato enters on transparent panel 21 in the input port 5 by upper box 10, rapid detent to position to be detected； Before potato starts detection, gear pushes away stepper motor 23 under the instruction of control module, and gear pushes away the driving gear push rod of stepper motor 23 26 return to reset condition, arrange in " two " font, take pictures convenient for video camera 27；

After potato enters on transparent panel 21, due to the effect of gravity, transparent panel 21 is moved downward, and four struts 13 push Spring 12, the rack structure 15 on strut 13 drive first gear 16 to rotate, and the bull stick 17 coaxial with first gear 16 rotates, and turns The rotation of bar 17 drives reflective mirror 20 to rotate, and the amplitude that the big potato of quality moves down transparent panel 21 is big, the small potato of quality The amplitude for moving down transparent panel 21 is small；So that reflective mirror 20 carries out angle adjustment according to the gravity size of potato, favorably Four virtual images of the potato lower part of four 20 internal reflections of reflective mirror are captured in video camera 27；

In order to obtain the characteristics of image of potato bottom to be checked, two sets of (staggered relatively) reflective mirrors 20 of configuration can be completed actual situation at As the panorama potato defects detection combined still for large-sized potato, exists because device space is too compact, no Large scale reflective mirror 20 can be placed, therefore, in addition configures two sets of (staggered relatively) reflective mirrors 20, wherein reflective mirror 20 can be by plane Mirror is changed to convex mirror, and Fig. 8 is plane mirror and convex mirror virtual image optical imagery schematic diagram；Thus, due to blocking for potato to be checked, It is unable to the potato of perfect imaging in plane mirror, the adaptation range of present apparatus detection can be promoted in convex mirror perfect imaging.

(3), the actual situation image checking stage；It first carries out taking pictures for the first time, obtains the image of part on potato to be detected, it should Image is real image, and the embedded-type ARM single-chip microcontroller in control module issues the annular LED lamp that the inner top of upper box 10 is opened in order 28, the lower LED light of 8 bottom of lower box is closed, after the top real image for obtaining potato to be measured, using image segmentation submodule to this Real image is handled, and the real image edge of potato is extracted, and is split, zooms to the fixed dimension of 256 × 256 pixels；It connects Taken pictures for the second time, embedded-type ARM single-chip microcontroller in control module issues the lower LED light that 8 bottom of lower box is opened in order, The annular LED lamp 28 at the inner top of upper box 10 is closed, the image of potato lower part is reflected in four reflective mirrors 20, which is The virtual image after video camera 27 obtains four virtual images of the lower part of potato to be measured, extracts four virtual image edges of potato respectively, and It is split, zooms to the fixed dimension of 128 × 128 pixels；Image segmentation submodule again will treated real image and four A virtual images are transported to the multichannel convolutive network submodular simultaneously and are identified, detection potato whether have germination, The defect of shagreen, breakage or deformity, and control module is sent by Industrial Ethernet by inspection result；

(4), darkroom detection device and sorting stage are released；After the completion of detection, gear pushes away stepper motor 23 in the instruction of control module Under, gear pushes away stepper motor 23 and rotates backward two gear push rods 26 of driving into " V " font arrangement, and potato is released upper box backward Delivery outlet 6 on 10, potato is moved on conveying body；For flawless potato in the defeated of conveying body It send down, moves directly to next process and handled, for defective potato, high pressure nozzle is opened in control module instruction Electric control valve in mechanism, high pressure nozzle mechanism spray high pressure draught and potato are pushed to the side of conveying body, from lacking Potato lower skateboard 3 is fallen into slide into defect potato collecting box 4.

Display sub-module in step (3) in control module is according to detection as a result, showing normal horse of current operation phase The number of bell potato, the number of all kinds of defect potatos, and detection sum；When device has it is abnormal when, pass through light warning submodule Block is warned；

As shown in figure 9, image preprocessing submodule, image segmentation submodule in step (3) intelligent algorithm defect recognition module Specific work process with multichannel convolutive network submodular is；

Top potato real image after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel convolutive A1 convolutional layer in network submodular, after the operation of 5 × 5 window convolutions, 15 width images of 254 × 254 pixels of generation, then by The pond A2 layer in multichannel convolutive network submodular carries out compression processing, generates 15 width images of 127 × 127 pixels, later, Second of convolution operation is carried out, the A3 convolutional layer being sent into multichannel convolutive network submodular is operated using 3 × 3 window convolutions Afterwards, 60 width images of 125 × 125 pixels are generated, then compression is carried out by the pond the A4 layer in multichannel convolutive network submodular Reason generates 60 width images of 63 × 63 pixels, further, at the full articulamentum of A5 in multichannel convolutive network submodular Reason exports the vector of 4096 dimensions, further, handles by the full articulamentum of A6 in multichannel convolutive network submodular, defeated The vector of 1024 dimensions out；

Lower part the first potato virtual image after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel B1 convolutional layer in convolutional network submodule generates 15 width images of 126 × 126 pixels after the operation of 5 × 5 window convolutions, Compression processing is carried out by the pond the B2 layer in multichannel convolutive network submodular again, generates 15 width images of 63 × 63 pixels, it Afterwards, second of convolution operation is carried out, the B3 convolutional layer being sent into multichannel convolutive network submodular is grasped using 3 × 3 window convolutions After work, 60 width images of 61 × 61 pixels are generated, then compression is carried out by the pond the B4 layer in multichannel convolutive network submodular Reason generates 60 width images of 31 × 31 pixels, further, at the full articulamentum of B5 in multichannel convolutive network submodular Reason exports the vector of 1024 dimensions, further, handles by the full articulamentum of B6 in multichannel convolutive network submodular, defeated The vector of 256 dimensions out；

Lower part the second potato virtual image after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel C1 convolutional layer in convolutional network submodule generates 15 width images of 126 × 126 pixels after the operation of 5 × 5 window convolutions, Compression processing is carried out by the pond the C2 layer in multichannel convolutive network submodular again, generates 15 width images of 63 × 63 pixels, it Afterwards, second of convolution operation is carried out, the C3 convolutional layer being sent into multichannel convolutive network submodular is grasped using 3 × 3 window convolutions After work, 60 width images of 61 × 61 pixels are generated, then compression is carried out by the pond the C4 layer in multichannel convolutive network submodular Reason generates 60 width images of 31 × 31 pixels, further, at the full articulamentum of C5 in multichannel convolutive network submodular Reason exports the vector of 1024 dimensions, further, handles by the full articulamentum of C6 in multichannel convolutive network submodular, defeated The vector of 256 dimensions out；

The lower part third potato virtual image after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel D1 convolutional layer in convolutional network submodule generates 15 width images of 126 × 126 pixels after the operation of 5 × 5 window convolutions, Compression processing is carried out by the pond the D2 layer in multichannel convolutive network submodular again, generates 15 width images of 63 × 63 pixels, it Afterwards, second of convolution operation is carried out, the D3 convolutional layer being sent into multichannel convolutive network submodular is grasped using 3 × 3 window convolutions After work, 60 width images of 61 × 61 pixels are generated, then compression is carried out by the pond the D4 layer in multichannel convolutive network submodular Reason generates 60 width images of 31 × 31 pixels, further, at the full articulamentum of D5 in multichannel convolutive network submodular Reason exports the vector of 1024 dimensions, further, handles by the full articulamentum of D6 in multichannel convolutive network submodular, defeated The vector of 256 dimensions out；

The 4th potato virtual image of lower part after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel E1 convolutional layer in convolutional network submodule generates 15 width images of 126 × 126 pixels after the operation of 5 × 5 window convolutions, Compression processing is carried out by the pond the E2 layer in multichannel convolutive network submodular again, generates 15 width images of 63 × 63 pixels, it Afterwards, second of convolution operation is carried out, the E3 convolutional layer being sent into multichannel convolutive network submodular is grasped using 3 × 3 window convolutions After work, 60 width images of 61 × 61 pixels are generated, then compression is carried out by the pond the E4 layer in multichannel convolutive network submodular Reason generates 60 width images of 31 × 31 pixels, further, at the full articulamentum of E5 in multichannel convolutive network submodular Reason exports the vector of 1024 pixels, further, handles by the full articulamentum of E6 in multichannel convolutive network submodular, defeated The vector of 256 dimensions out；

Finally, exporting vector, the B6 of 1024 dimensions to the full articulamentum of A6 by the A7 tether layer in multichannel convolutive network submodular The full articulamentum of vector, D6 that the full articulamentum of vector, C6 that full articulamentum exports 256 dimensions exports 256 dimensions exports 256 dimensions The vector that the full articulamentum of vector sum E6 exports 256 dimensions executes set add operation, generates lacking for characterization potato panoramic surface Fall into characteristic information, be the vector of 2048 dimensions, then from soft the recurrences layer of the A8 in multichannel convolutive network submodular export 5 tie up to Amount indicates that potato to be detected belongs to the probability density distribution of normal, germination, shagreen, breakage or lopsided surface defect, to distinguish Potato to be checked is known with the presence or absence of defect；

Multichannel convolutive network submodular includes 5 channels, can disposably input 1 potato upper part real image image to be detected With part virtual images under 4 width；The off-line training sample database of multichannel convolutive network submodular can increase sample size；Cause And the germination of potato, shagreen, breakage, lopsided type can further be segmented with the increase of sample size；Multichannel volume Product network submodular can be both trained update in use by user, also can choose regular by device manufacturer It updates；The present apparatus supports the multichannel convolutive network submodular of multi version, can be carried out by end user according to practical application scene Autonomous selection.

The present embodiment not makes any form of restriction shape of the invention, material, structure etc., all according to this hair Bright technical spirit any simple modification, equivalent change and modification to the above embodiments, belong to the technology of the present invention side The protection scope of case.

Claims (10)

1. the panorama potato defect detecting device combined is imaged based on actual situation, it is characterised in that: including conveying device, actual situation Imaging combines formula darkroom detection device and control cabinet；

Conveying device include feeding conveying mechanism, conveying body, high pressure draught nozzle mechanism, defect potato lower skateboard and Defect potato collecting box；Output end on rear side of feeding conveying mechanism combines on front side of the detection device of formula darkroom with actual situation imaging Input port connection, actual situation imaging combine the input terminal on front side of delivery outlet and conveying body on rear side of the detection device of formula darkroom Connection；The left side of conveying body is arranged in high pressure draught nozzle mechanism, and defect potato lower skateboard is located at outfeed belt conveyor The right side of structure is simultaneously oppositely arranged with high pressure draught nozzle mechanism or so, and defect potato lower skateboard is obliquely installed in right low left high, The lower section of defect potato lower skateboard right edge is arranged in defect potato collecting box；The rear side output end of conveying body is Qualified potato outlet；

Be provided with control module and intelligent algorithm defect recognition module in control cabinet, control module by Industrial Ethernet respectively with Conveying device combines formula darkroom detection device with actual situation imaging and connects, and intelligent algorithm defect recognition module and control module connect It connects.

2. according to claim 1 the panorama potato defect detecting device combined is imaged based on actual situation, feature exists In: it includes vertically disposed four columns that actual situation imaging, which combines formula darkroom detection device, and top is provided between four columns Open lower box, the upper end of each column has been bolted a connector, has been provided with bottom between four connectors The upper box of portion's opening, upper box and lower box are oppositely arranged up and down and have gap, and actual situation imaging combines the detection of formula darkroom The delivery outlet of input port and rear side on front side of device is separately positioned on the lower front and back lower of upper box, input port and defeated Export correspondence setting；

There are four the branch cover for seat of open top, four branch cover for seat are respectively adjacent to four sides of lower box for bottom setting in lower box Wall, the line of four branch cover for seat are diamond shape, and a spring is provided in each cover for seat, and spring upper end is higher than branch cover for seat, often It is inserted with the strut being arranged in vertical in a spring, the gear with spring upper end press cooperation is provided on strut Ring, strut lower end are extend into branch cover for seat and apart from branch cover for seat bottom 2-5cm, the right side of the lower part of left side strut, right side strut The left side of lower part, be provided with rack structure on front side of the rear side of lower part of front side strut and the lower part of rear side strut, often A rack structure is engaged with a first gear, the central coaxial of first gear to being fixedly installed a bull stick, bull stick Both ends are rotatably connected on the opposite two sides inner wall of lower box by bearing, and the medium position of bull stick is arranged reflective by mounting rack The reflective mirror that face tilts upward；The upper end of four struts is horizontally disposed one piece of transparent panel, and transparent panel is in rectangle and and lower box There is the gap 1-3mm between four side walls；It is respectively arranged with a potato gear on the strut on left side and strut and right side and pushes away machine Structure, two potato gear pushing mechanism note constructions are identical and are symmetrical set.

3. according to claim 2 the panorama potato defect detecting device combined is imaged based on actual situation, feature exists In: the potato gear pushing mechanism positioned at right side includes support plate, and support plate is fixedly connected in lower box outer right wall, support plate Lower surface be provided with gear and push away stepper motor, the main shaft that gear pushes away stepper motor passes through support plate vertically upward and is coaxially connected with the Two gears, support plate upper surface are rotatably connected to third gear, and second gear and third gear are respectively positioned on the right edge of transparent panel The right side on edge, third gear are engaged with second gear, and the center of third gear upper surface is fixedly connected with horizontally disposed gear and pushes away Bar, the lower surface for keeping off push rod are higher than transparent panel upper surface 1-3mm, keep off central symmetry of the both ends of push rod about third gear； The gear push rod of two potato gear pushing mechanisms has when turning in straight line, between two gear push rods between 2-10mm Gap, gap of the gear push rod in rotation across upper box and lower box.

4. according to claim 2 the panorama potato defect detecting device combined is imaged based on actual situation, feature exists In: video camera is installed, the periphery of video camera is provided with annular LED lamp at the center of upper box inner top；Bottom in lower box Portion is provided with the lower LED light irradiated upward.

5. according to claim 1 the panorama potato defect detecting device combined is imaged based on actual situation, feature exists It include that multiple groups are successive or one group of belt conveyor in: feeding conveying mechanism and conveying body, Belt Conveying Mechanism includes bracket, and bracket left or right side is provided with decelerating motor, and drive roll and driven voller are rotatably connected on bracket, actively Roller and driven voller are horizontally disposed in left-right direction, pass through the belt rotation connection of annular, belt between drive roll and driven voller It is in one that several are arranged at intervals in the middle part of outer surface for accommodating the groove of single potato, groove, input port and delivery outlet The setting of straight line, the output end of decelerating motor are provided with driving pulley, and one end of drive roll is provided with driven pulley, driving pulley It is connected with driven pulley by V V belt translation.

6. according to claim 5 the panorama potato defect detecting device combined is imaged based on actual situation, feature exists In: control module includes embedded-type ARM single-chip microcontroller, I/O submodule, Industrial Ethernet communication submodule, display sub-module, light Warn submodule and computer；Computer be intelligent algorithm defect recognition module operation physical support, computer and industry with Too Network Communication submodule is connected；Embedded-type ARM single-chip microcontroller communicates submodule, I/O submodule, display submodule with Industrial Ethernet Block is connected with light warning submodule, this is the key submodule of control module；I/O submodule respectively with decelerating motor, gear It pushes away step motor control signal to be connected, corresponding gear is successively controlled according to control flow pushes away stepper motor and orderly act, institute It states I/O submodule to be also connected with the electric control valve in high pressure draught nozzle mechanism, according to recognition result, turns on or off electricity Air control valve；The I/O submodule is also connected with and controls the opening and closing of annular LED light and lower LED light；The computer is total by USB Line mode is connected with video camera, and by Industrial Ethernet communication mode, computer is pressed the instruction of embedded-type ARM single-chip microcontroller, opened Dynamic video camera is taken pictures, and emphasis of taking pictures for the first time obtains parts of images on potato, and emphasis of taking pictures for the second time obtains under potato Parts of images, and photo is deposited into computer.

7. according to claim 6 the panorama potato defect detecting device combined is imaged based on actual situation, feature exists In: intelligent algorithm defect recognition module includes image preprocessing submodule, image segmentation submodule and multichannel convolutive network Module composition；Described image pretreatment submodule pre-processes the image taken pictures twice, promotes the quality of image, has handled Image afterwards is sent to image segmentation submodule；Described image segmentation submodule handles image of taking pictures for the first time, obtains The real image edge of potato, and it is split, zooms to the fixed dimension of 256 × 256 pixels, described image divides submodule pair Image of taking pictures for the second time is handled, and four virtual image edges of potato are obtained, and is split, is zoomed to 128 × 128 pixels Fixed dimension, a real image and four virtual images are transported to the multichannel again simultaneously and rolled up by described image segmentation submodule Product network submodular is identified, detects potato sprouting, shagreen, breakage or a series of defects of deformity, and inspection result is led to It crosses Industrial Ethernet and is sent to control module.

8. using the detection side of the panorama potato defect detecting device combined as claimed in claim 7 based on actual situation imaging Method, it is characterised in that: include the following steps,

(1), into the preparation stage before the detection device of darkroom；The start and stop of decelerating motor, feeding conveying are controlled by control module Ensure that only one potato to be transported occurs in each groove on the belt of mechanism, prevents multiple potatos to be checked simultaneously Enter darkroom detection device；

(2), the darkroom detection device detent stage；Before entering darkroom detection device, the gear of potato gear pushing mechanism pushes away potato Under the instruction of control module, gear pushes away stepper motor driving gear push rod and arranges in front wide and rear narrow "eight" shape stepper motor, makes After potato enters on transparent panel in the input port by upper box, rapid detent to position to be detected；When potato opens Begin before detection, gear pushes away stepper motor under the instruction of control module, and gear pushes away stepper motor driving gear push rod and returns to reset condition, is in " two " font arrangement, takes pictures convenient for video camera；

After potato enters on transparent panel, due to the effect of gravity, transparent panel is moved downward, pressing spring under four struts, Rack structure driving first gear rotation on strut, the coaxial bar rotation with first gear, bar rotation drive reflective mirror Rotation, the amplitude that the big potato of quality moves down transparent panel is big, and the amplitude that the small potato of quality moves down transparent panel is small；From And reflective mirror is made to carry out angle adjustment according to the gravity size of potato, be conducive to video camera and capture four reflective mirror internal reflections Potato lower part four virtual images；

(3), the actual situation image checking stage；It first carries out taking pictures for the first time, obtains the image of part on potato to be detected, the image For real image, the embedded-type ARM single-chip microcontroller in control module issues the annular LED lamp that top in upper box is opened in order, under closing The lower LED light of the bottom of box, after the top real image for obtaining potato to be measured, using image segmentation submodule to the real image at Reason, extracts the real image edge of potato, and be split, zoom to the fixed dimension of 256 × 256 pixels；Then second is carried out Secondary to take pictures, the embedded-type ARM single-chip microcontroller in control module issues the lower LED light that lower box bottom is opened in order, closes upper box In top annular LED lamp, the image of potato lower part is reflected in four reflective mirrors, which is the virtual image, and video camera obtains In reflective mirror to be measured after four virtual images of the lower part of potato, four virtual image edges of potato are extracted respectively, and be split, Zoom to the fixed dimension of 128 × 128 pixels；Image segmentation submodule again will a treated real image and four virtual images Be transported to the multichannel convolutive network submodular simultaneously and identified, detection potato whether have germination, shagreen, breakage or The defect of deformity, and control module is sent by Industrial Ethernet by inspection result；

(4), darkroom detection device and sorting stage are released；After the completion of detection, gear pushes away stepper motor under the instruction of control module, Gear pushes away stepper motor and rotates backward two gear push rods of driving into " V " font arrangement, potato is released backward defeated on upper box Outlet, potato are moved on conveying body；For flawless potato under the conveying of conveying body, directly It is moved to next process to be handled, for defective potato, control module instruction is opened in high pressure nozzle mechanism Electric control valve, high pressure nozzle mechanism sprays the side that potato is pushed to conveying body by high pressure draught, from defect potato Lower skateboard slides into defect potato collecting box.

9. detection device method according to claim 8, it is characterised in that: display in step (3) in control module Module according to detection as a result, show the number of current operation phase Normal potato, the number of all kinds of defect potatos, and Detection sum；When device has it is abnormal when, warned by light warning submodule.

10. detection device method according to claim 8, it is characterised in that: step (3) intelligent algorithm defect recognition module In image preprocessing submodule, image segmentation submodule and multichannel convolutive network submodular specific work process be；

Top potato real image after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel convolutive A1 convolutional layer in network submodular, after the operation of 5 × 5 window convolutions, 15 width images of 254 × 254 pixels of generation, then by The pond A2 layer in multichannel convolutive network submodular carries out compression processing, generates 15 width images of 127 × 127 pixels, later, Second of convolution operation is carried out, the A3 convolutional layer being sent into multichannel convolutive network submodular is operated using 3 × 3 window convolutions Afterwards, 60 width images of 125 × 125 pixels are generated, then compression is carried out by the pond the A4 layer in multichannel convolutive network submodular Reason generates 60 width images of 63 × 63 pixels, further, at the full articulamentum of A5 in multichannel convolutive network submodular Reason exports the vector of 4096 dimensions, further, handles by the full articulamentum of A6 in multichannel convolutive network submodular, defeated The vector of 1024 dimensions out；

Lower part the first potato virtual image after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel B1 convolutional layer in convolutional network submodule generates 15 width images of 126 × 126 pixels after the operation of 5 × 5 window convolutions, Compression processing is carried out by the pond the B2 layer in multichannel convolutive network submodular again, generates 15 width images of 63 × 63 pixels, it Afterwards, second of convolution operation is carried out, the B3 convolutional layer being sent into multichannel convolutive network submodular is grasped using 3 × 3 window convolutions After work, 60 width images of 61 × 61 pixels are generated, then compression is carried out by the pond the B4 layer in multichannel convolutive network submodular Reason generates 60 width images of 31 × 31 pixels, further, at the full articulamentum of B5 in multichannel convolutive network submodular Reason exports the vector of 1024 dimensions, further, handles by the full articulamentum of B6 in multichannel convolutive network submodular, defeated The vector of 256 dimensions out；

Lower part the second potato virtual image after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel C1 convolutional layer in convolutional network submodule generates 15 width images of 126 × 126 pixels after the operation of 5 × 5 window convolutions, Compression processing is carried out by the pond the C2 layer in multichannel convolutive network submodular again, generates 15 width images of 63 × 63 pixels, it Afterwards, second of convolution operation is carried out, the C3 convolutional layer being sent into multichannel convolutive network submodular is grasped using 3 × 3 window convolutions After work, 60 width images of 61 × 61 pixels are generated, then compression is carried out by the pond the C4 layer in multichannel convolutive network submodular Reason generates 60 width images of 31 × 31 pixels, further, at the full articulamentum of C5 in multichannel convolutive network submodular Reason exports the vector of 1024 dimensions, further, handles by the full articulamentum of C6 in multichannel convolutive network submodular, defeated The vector of 256 dimensions out；

The lower part third potato virtual image after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel D1 convolutional layer in convolutional network submodule generates 15 width images of 126 × 126 pixels after the operation of 5 × 5 window convolutions, Compression processing is carried out by the pond the D2 layer in multichannel convolutive network submodular again, generates 15 width images of 63 × 63 pixels, it Afterwards, second of convolution operation is carried out, the D3 convolutional layer being sent into multichannel convolutive network submodular is grasped using 3 × 3 window convolutions After work, 60 width images of 61 × 61 pixels are generated, then compression is carried out by the pond the D4 layer in multichannel convolutive network submodular Reason generates 60 width images of 31 × 31 pixels, further, at the full articulamentum of D5 in multichannel convolutive network submodular Reason exports the vector of 1024 dimensions, further, handles by the full articulamentum of D6 in multichannel convolutive network submodular, defeated The vector of 256 dimensions out；

The 4th potato virtual image of lower part after image segmentation submodule is converted to the monochromatic 3 width images of RGB, is sent into multichannel E1 convolutional layer in convolutional network submodule generates 15 width images of 126 × 126 pixels after the operation of 5 × 5 window convolutions, Compression processing is carried out by the pond the E2 layer in multichannel convolutive network submodular again, generates 15 width images of 63 × 63 pixels, it Afterwards, second of convolution operation is carried out, the E3 convolutional layer being sent into multichannel convolutive network submodular is grasped using 3 × 3 window convolutions After work, 60 width images of 61 × 61 pixels are generated, then compression is carried out by the pond the E4 layer in multichannel convolutive network submodular Reason generates 60 width images of 31 × 31 pixels, further, at the full articulamentum of E5 in multichannel convolutive network submodular Reason exports the vector of 1024 pixels, further, handles by the full articulamentum of E6 in multichannel convolutive network submodular, defeated The vector of 256 dimensions out；

Finally, exporting vector, the B6 of 1024 dimensions to the full articulamentum of A6 by the A7 tether layer in multichannel convolutive network submodular The full articulamentum of vector, D6 that the full articulamentum of vector, C6 that full articulamentum exports 256 dimensions exports 256 dimensions exports 256 dimensions The vector that the full articulamentum of vector sum E6 exports 256 dimensions executes set add operation, generates lacking for characterization potato panoramic surface Fall into characteristic information, be the vector of 2048 dimensions, then from soft the recurrences layer of the A8 in multichannel convolutive network submodular export 5 tie up to Amount indicates that potato to be detected belongs to the probability density distribution of normal, germination, shagreen, breakage or lopsided surface defect, to distinguish Potato to be checked is known with the presence or absence of defect；

Multichannel convolutive network submodular includes 5 channels, can disposably input 1 potato upper part real image image to be detected With part virtual images under 4 width；The off-line training sample database of multichannel convolutive network submodular can increase sample size；Cause And the germination of potato, shagreen, breakage, lopsided type can further be segmented with the increase of sample size；Multichannel volume Product network submodular can be both trained update in use by user, also can choose regular by device manufacturer It updates；The present apparatus supports the multichannel convolutive network submodular of multi version, can be carried out by end user according to practical application scene Autonomous selection.