CN106127749A - The target part recognition methods of view-based access control model attention mechanism - Google Patents

Abstract

The invention discloses the target part recognition methods of a kind of view-based access control model attention mechanism, including: vision noticing mechanism Model Selection, select the attention mechanism model of feature based and attention mechanism model based on space, utilize biological central peripheral filter construction, multiple space scales extract color, direction, brightness；What feature combined significantly schemes to generate, and color, direction and brightness is combined into characteristic pattern, thus the significance obtaining correspondence describes, and will be formed and significantly scheme after normalization calculating and linear combination；Target manufactured parts recognition strategy, carries out Parts Recognition according to collection part image, the notable figure of generation, binary conversion treatment and optimization, extraction marking area, the flow process of part zone extraction.This invention enable the visual system of industrial robot effectively to identify work space in target part, be accurately positioned target part, make industrial robot have higher autonomy, robustness, adaptability when completing manufactured parts assembling work.

Description

The target part recognition methods of view-based access control model attention mechanism

Technical field

The present invention relates to the technical field of Parts Recognition, the target part particularly to a kind of view-based access control model attention mechanism is known Other method.

Background technology

Carry out Parts Recognition and location by machine vision, and guide the mechanical hand of industrial robot to capture and Assembly part It it is the key issue in industrial robot application.At present, on all kinds of flow production lines, industrial robot performs at most Operation exactly target part is identified, positions, captures, installs.But sometimes due to affected by many objective factors, Position and the attitude of target part are it may happen that change, if industrial robot goes to know also according to pre-set program , do not position, capture, if installation targets part, it is likely that many beyond thought consequences can occur.Therefore, in order to improve The adaptation ability of flow production line, needs industrial robot that the target part captured and install is identified efficiently and determined Position, obtains the three-dimensional position of target part and attitude, thus the mechanical hand controlling industrial robot goes to capture accurately and Install.Additionally, developing rapidly along with small lot and multi-item production mode, for variety classeses many on flow production line and For the part of profile and size, method the most manually is identified and positions and cannot efficiently and accurately complete to make at all Industry task, needs to use machine vision technique and method can quickly complete target part identification, so that part assembling Operation has more preferable motility, robustness, high efficiency etc..

Machine vision technique is to grow up the sixties in 20th century.Massachusetts Institute Technology's Lincoln laboratory in 1961 Use photographic head as the input of computer, object identification and image processing method are incorporated in robot application.Thus open The machine vision that begun research.As the well-known industrial robot enterprise in the whole world, the Adept Technology company of the U.S. is grinding Machine vision technique is just added during first-generation industrial robot processed.In R&D process subsequently, the said firm obtains abundant Machine vision experience and ripe machine vision technique, this has promoted it as the maximum industrial robot of the U.S. and has manufactured public affairs Department.

(1) vision noticing mechanism general introduction

The mankind are recognized and the perception world by vision, audition, olfactory sensation, sense of touch, and wherein vision is again that the mankind obtain environment The most important approach of information, so the research to this respect always is the focus that each side pays close attention to.But, when people enter one In visual scene, the visual information blown against one's face is magnanimity (per second about 10⁸～10⁹Bit), but human visual system is to letter The ability that breath processes but is limited, it is impossible to meet the needs processing all data in real time.At this moment, Primates class biology is possessed The ability of superpower real-time process complex environment data attract attention.This ability can process further Before visual information, selectively input sub-fraction data therein and carry out point as the data of " interested " and " meaningful " Analysis and process.The most one by one it is switched to different focus, thus realizes the very big reduction of amount of calculation, improve simultaneously The efficiency of Vision information processing.This ability of rapid extraction critical data in mass data of Primates class animal, is claimed For Visual Selective Attention mechanism.Being the example of several vision attention as shown in Figure 2, as in Fig. 2 (a), empty circles can head First it is noted；In Fig. 2 (b), then it is to it is initially noted that solid circles, and ignores other square.

Vision noticing mechanism research is started from neurobiology and psychological field at first, is used for studying some mental models And cognitive model.Along with image procossing and the development of Vision Builder for Automated Inspection, increasing attention mechanism model is in these fields Interior generation and development.Several applications of vision noticing mechanism are: 1. target detection and target recognition.In the face of one complicated During visual scene, the existence of other object can have a strong impact on the detection to target object and identification, the most topmost shadow in a large number Ring or the efficiency of image procossing.Because substantial amounts of calculating resource all spends in the identification of other object, the quantity of information of process The most relatively large with the complexity calculated.The effect introducing vision noticing mechanism is exactly reasonable distribution can to calculate resource, at general The center of gravity of reason is placed on the object of those " doubtful " or " concern ".2. the compression of image and coding.More existing compare main flow Method for compressing image, the overwhelming majority is all data of image to be put on an equal footing, and uses same Compression Strategies and compression Ratio is compressed, and this is for each specific application scenarios, a wise move the most at last.Because regarding for people For the visual processes of vision system or computer, each application scenarios simply need a part of data in image carry out Process, such as: pedestrian's monitoring concern is primarily with people, and other information being in picture is dispensable；Vehicle monitoring is then Paying close attention to car, other information does not has the biggest meaning.Generally speaking, introduce vision noticing mechanism can be extracted by key message Come, take the Compression Strategies of some high-fidelities, for the information that other is secondary, then can take some lossy compression method strategies.This Sample just can be while ensureing higher data compression ratio, moreover it is possible to ensures that the key message in image is not lost.3. image inspection Rope and image classification.The research direction that field of image search is main now is content-based retrieval mode, how to extract point Distinguish that characteristics of image in hgher efficiency is to problem to be solved in representing the content of correspondence image and being exactly this field.And note Mechanism may be exactly an effective and feasible solution, extracts the key message in image in characterizing whole image Hold, obtain the similarity between image by contrast key message.As such, it is possible to improve image retrieval and the efficiency of classification.With Time, additionally it is possible to effectively reduce or avoid the impact of uncorrelated content in image.

(1) vision noticing mechanism is theoretical

1) feature integration is theoretical (Feature Integrated Theory).Feature integration theory is by brother's rival British Sub-university Treisman professor and Oxford University professor Gelade propose, and receive the extensive pass in attention mechanism field afterwards Note, and a large amount of vision noticing mechanism computation models of formation and development on this basis.Feature integration for attention mechanism is managed Opinion, it should be noted that single holistic object mainly has two ways: noted by focus or believed by vision from top to down Breath processes.But be in certain set environment, it is impossible to determine be which kind of mode in action, or do not know which kind of mode More useful to visual system.Under normal circumstances, both modes are collaborative works, the most under some extreme conditions, can see Can be operated in the most completely self-contained mode to the two mode.The structure of this theory is as it is shown on figure 3, describe vision letter The basic handling flow process of breath and the process of the generation of region-of-interest.

The object identification mode noted based on focus is referred to as feature registration phase by Treisman, notes the stage the most in advance: this Time visual system whole image-region is gathered concurrently in the way of " luminous point " low-level image feature (include color, direction, size, Distance etc.), for the mankind, almost it is unaware of this full automatic processing procedure.The pre-process noting the stage People can be promoted to carry out referring to the exploration of aeoplotropism in visual scene.But, vision commitment can only detect single independence Feature, and the relation between each feature can not be detected, contacting between feature and position can not be determined.Then to spy Levy employing corresponding template and carry out independent coding, generate characteristics map FM (Feature Map).As the color in Fig. 3, direction, Size etc..Each multidimensional characteristic comprises multiple characteristic pattern, such as color characteristic then can comprise three kinds of characteristic patterns of red, green, blue.So Rear feature based figure sets up the location drawing (Map ofLocation) of reflection saliency, and this may indicate that concern object is at figure Location in Xiang, but cannot differentiate what the object in marking area is.Treisman by another kind based on from upper and Lower process object recognition method is referred to as the feature integration stage, starts to identify object in this stage.When object blocks, just Focus can be had a strong impact on and note process, so needing to identify object by top-down process.Now, visual system by Marking area on individual scan position figure, is combined correlated characteristic in region according to set mode, to generate a certain The sign of object.Visual system, when processing position, needs to focus on original, separate feature integration being one Individual single object.When attention is disperseed or exceedes people's ability to bear, arise that the most proper for some stimulation characteristic Local combination, i.e. illusory conjunction (Illusory Conjunctions), thus cause illusion phenomenon.

The establishment of feature integration theory imposes a condition based on following: notice that the stage is in image, some are simple and can use the most in advance Data encode, generate some characteristic modules.These modules may have the information about spatial relation, but they Directly can not provide the data of these mark locus to the feature integration stage.2. when centrality notes generation effect, just Can start in situational map, the feature being in marking area is extracted and is combined.By the process of attention mechanism, currently During in selected localization, all features all can be fused to an interim object sign or file.3. the visual system knowledge to object Other process is: first integrates in each object file about its character and the information of structural relation, then forms the description letter of correspondence Breath.Finally this being described and identify that in network (Recognition Network), the Object representation of storage compares, coupling is then Represent and successfully identify object.There is following character in feature integration theory: is familiar with in scene at one, and some predictable objects can To determine, without going again to check they relations in space by mating their some independent characteristics.The most right The search of target object is the most relatively easy, and detection efficiency is the highest.But be in one and be not familiar with in scene, or task is correlated with Connection be some union features (Conjunctive Features) time, to target object search efficiency will reduce.Such as During search face, when this person is in countryside tourism, can quickly find out；But when he and companion wear unified clothing according to collection When body shines, him is found to be accomplished by spending the biggest energy, even if this person is very familiar with.Feature analysis is characteristic theory with identification network The key of middle identification target, new perception can occur when file destination is replaced.

2) guiding search is theoretical (Guided Search Theory).Guiding search theory is Harvard University Wolfe professor Within 1994, proposing, this theoretical description neuromechanism of vision attention, the development for cognitive neuropsychology provides theory Basis.Guiding search theory is initially on the basis of feature integration theory does some adjustment, and then combining Neisser will regard Feel processes and is divided into the pre-attention stage and notices that the theory in stage, the two-stage model of Hoffman and Egeth are in characteristic binding side The achievement in research in face.The main purpose of guiding search model is used to explain that people are one normal, in continuous print visual scene Find out the ability in the visual stimulus source of needs, it considers that the significance that the mankind mainly visually stimulate according to external object selects Select the object of attention.It is the same with feature integration theory all thinks that single features search all comprises two rank with union feature search Section: parallel and sequence stage, but the former is the basis of the latter.Its model framework as shown in Figure 4, can according to this model framework Know that its workflow is as follows: first, parallelization process whole image, allowed stimulus penetrating cross broad tuning classification passage, generate Limited basic visual signature group, i.e. characteristics map FM.Relatively independent spy can be had for color, direction and size etc. Expropriation of land figure, can also be refined in multiple independent characteristics map, this point and feature integration theory in each characteristic type Characteristics map be similar to.Additionally, all features can centralized displaying in a single multidimensional characteristic map.Generating spy After expropriation of land figure, system just determines significance degree by the activation degree calculating regional.For certain specific region, it swashs The degree of living is the biggest, it is meant that the probability directly noted is the biggest.The calculating of activation degree is made up of two parts: stimulates and is derived from The bottom-up type of body character-driven activates the top-down type activation degree that degree drives with user task.The former is used for marking Knowing a certain object particularity degree in current scene, the latter is for extracting the object that those tasks need, and is typically spy The object that different property is relatively low.Activation degree for both types also has two calculating principles: be correlated with for certain Top-down information characteristics, then the weight arranging its top-down activation degree is 0；Certain attribute and target when objects interfered When object is identical, then the bottom-up reducing this feature activates degree weight.Finally, by calculated to regional Activation degree is weighted sum operation, thus generates activation map AM (Activation Map), and visual system is the most in order The process resource of distribution limited capacity.And can be by each characteristic model as being some topography, mountain peak therein is exactly Those activate the region that degree is higher, and attention then concentrates on these mountain peaks.But, the high activation degree that mountain peak is identified Be not offered as its information having any mark object, its Main Function be only intended to guide attention, do not have any in During order, will be it is first noted that the region of the highest activation degree.In visual search task, if not finding in current region Object, then can reconfigure attention region and have on the region of the highest activation degree to the next one, by that analogy until finding Target or appearance are searched for unsuccessfully.

3) competition theory (Integrated Competition Theory) is integrated.Integrating competition theory is with Britain's sword Bridge university professor Duncan proposes for representative.Vision attention theoretical definition is found many for research people's reply by Duncan simultaneously The method that the ability of individual object limits.Meanwhile, these theories are divided into three big factions by him: based on object, difference and space Vision attention is theoretical.They are paid close attention to the restriction of the standalone object number simultaneously discovered respectively, can generate independent difference domain number Purpose limits and can extract the restriction of information in area of space.Integrating competition theory is then object-based vision attention One of important foundation of Mechanism Model research.Integrate competition theory and derive from the visual attention model in neuropsychology, at this In model, visual information is embodied on people's brain system of multiple eye response for the competition noted, including sensation and kinetic system System, cerebral cortex and hypodermic layer system.The integration competed by multisystem, can select same object as noting target.Whole The establishment closing competition theory depends on following general principle: first, the input of visual information can cause multiple brain system Participate in, here it is considered that in most systems, the processing procedure of visual information is vied each other: be right to certain when strengthening During the response of elephant, may result in and weaken the response to other object.Secondly, the top-down information startup to neuronal activity Can affect and the competition of current behavior related object.Such as, when performing to search green character task, to face in texturing systems Color encodes, then selecting the neuron of green visual information to pre-actuate, other color will be by greatly simultaneously Suppression, thus cause the display of green character can occupy competitive advantage.Finally, although competition occurs in each brain system, But the sensorimotor neutral net eventually by each several part completes the integration of competition information.If an object at some is Obtaining in system and pay close attention to, then in other systems, this object can obtain substantial amounts of response too and process resource.This information Integration be necessary time particular task is started from for target selection.

(3) vision noticing mechanism model

According to physiology and psychological field achievement in research, note being divided into attention based on space and object-based attention Two kinds.Classify according to the two, produce on the basis of above classical theory and developed following attention mechanism computation model.

1) attention mechanism model based on space

The most bottom-up (bottom-up) attention model.Bottom-up attention is a kind of in image basis, by data Drive the mechanism of focus.The feature integration theory of Treisman is the basis of bottom-up attention model, also by reality Checking understands that the attention of the type can work alone and effectiveness under some extreme condition.It addition, these bottom-up notes The realization of meaning model, an also important ingredient, it is simply that Koch calculates the framework of significance.Itti model is space note The classical representative of meaning model, it is that first complexity can calculate attention mechanism model efficiently.The general frame of Itti model is such as Shown in Fig. 5, it is as follows that it realizes process.First stage: early stage primary image feature extraction.Itti model uses linear filter to carry Take characteristics of image, and under each passage and yardstick, set up feature gaussian pyramid respectively.Then, on the basis of nonuniform sampling On by central peripheral poor (Center-surround Difference) computational methods extract characteristic pattern FM.With color characteristic figure As a example by calculating, if r, g, b represent the red, green, blue tristimulus values in cromogram respectively, then there are red R, green G, blue B, yellow tetra-widebands of Y Color Channel component is:

R=r-(g+b)/2, G=g-(r+b)/2, B=b-(r+g)/2, Y=(r+g)/2-r-g/2-b (1)

In formula, when numerical value is negative, then it is set to 0.Then, based on gaussian pyramid according to central peripheral difference Theoretical Calculation Red green two-way comparison diagram RG, and blue yellow two-way comparison diagram BY:

RG (c, s)=(R (c)-G (c)) Θ (G (s)-R (s), BY (c, s)=(B (c)-Y (c)) Θ (Y (s)-B (s) (2)

In formula, c and s all represents gaussian pyramid yardstick, and c ∈ 2,3,4}, s=c+ δ, δ ∈ 3,4}, represent this model The different scale selected.In like manner, the direction on brightness figure and four direction θ={ 0 °, 45 °, 90 °, 135 ° } can be obtained Characteristic pattern.

Second stage: generate notable figure (Saliency Map), uses Standardization Operator N () to carry out multiple features merging. All characteristic patterns are merged operation, generates notable figure.Then according to the focus of attention in notable figure, position or change pass Note region.In the case of Standardization Operator is only applicable to not have top-down information, its process processing each characteristic pattern is: 1. Value on standardized feature figure is in fixed range [0, M], to eliminate the amplitude difference between each feature；2. the overall situation is found out maximum Value M, calculates the average of other local maxima3. the last overall situation is multiplied byWith the prominent overall situation to the full extent maximum with Difference between local maximum.Calculate the notable figure of each feature dimensions according to Standardization Operator, yardstick is brightness when 4, face Color, the notable figure computational methods in direction are respectively as follows:

In formula,Represent across yardstick summation operation.Then, the notable figure after these three channel standard is directly closed And significantly scheme on S to final:

$S=\frac{1}{3}\left(N\right(\stackrel{\‾}{I})+N(\stackrel{\‾}{C})+N(\stackrel{\‾}{O}\left)\right)---\left(4\right)$

By the strategy that the victor is a king, the maximum region on notable figure is selected to pay close attention to as initial focus, switching Use was needed to forbid return mechanisms before next maximum, it is to avoid the same region of repeated accesses.Generally, Itti algorithm tool There are preferable robustness and accuracy.But, the 1/256 of the notable figure only artwork size of Itti model generation, this is great Limit its application.So, follow-up with Itti model for instructing, create some full-scale saliency extracted region and calculate Method.

The most top-down (top-down) attention model.Top-down attention is a kind of on the basis of study, by appointing Business or the mechanism of Knowledge driving focus of attention.Due on physiology or psychological study, for the specific works of attention mechanism Mode does not the most check on, and this causes almost without available top-down attention mechanism computation model.But, those leave The guide of upper layer data, relies only on attention mechanism computation model produced by the information on picture, can have at target self information Limit, position dimension variable and notable not time, cause its efficiency to be substantially reduced.In the research of top-down special aspect, relatively A prominent achievement is the task orientation model that Navalpakkam sets up.This model is the work at its teacher Itti On the basis of formed, its framework is as shown in Figure 6.The working mechanism of this model is: vision brain is the most meaningful to note in guiding figure Image block be current marking area.Act on behalf of the communication interface primarily as other three parts, working memory and long-term note Recall two knowledge bases for calculating the dependency of current marking area, and for the relevant figure of more new task.The circulation of target recognition End condition is, finds all related entities at task figure of being correlated with.

2) object-based attention Mechanism Model

Integrating on the basis of competition theory, attention mechanism model is to switch over for object with object, select focus, Thus generate object-based attention model.Meanwhile, these models are based on following viewpoint: pre-, visual system notices that the stage will Acquired image carries out preliminary clusters analysis, and attention mechanism just chooses specific objective in these cluster targets simultaneously.Sun To propose visual attention computation model of based on target the most wide-spread with Fisher, this model development Duncan's Integrate competition theory, and merge the vision significance computation model of Koch and Itti etc., bottom-up and top-down Between vision attention mutual, object-based attention is machine-processed and in the combination of attention mechanism based on space, target and between target The knowledge blocks such as visual representation.In this model, the competition of attention is not only present in target internal, there is also and each target Between, and the transfer of attention layering is carried out.This model framework is as shown in Figure 7.What Sun-Fisher model mainly solved asks Topic can be summarized as following some: first, principal character extract.This significance model proposed with Itti etc. is similar, will input Picture breakdown becomes a series of multiple dimensioned characteristic pattern, thus produces four color pyramids (R, G, B, Y), a brightness pyramid (I) and four direction pyramid (θ={ 0 °, 45 °, 90 °, 135 ° }), it is achieved a kind of bottom-up attention.Secondly, marshalling is aobvious Work maps.This module achieves object-based attention and object-based attention is merged, and defines a target and space Hierarchical structure.One marshalling can be a point, object, can also be the hierarchical structures of multiple marshalling composition.Marshalling Significance not only determined also have the shadow cooperated with each other and compete between the internal each parts of marshalling by environment about simultaneously Ring.End, it is noted that competition and focus shift.This operation is multiple dimensioned in a model, has by first all to the layer of thin yardstick Aggregated(particle) structure.Focus of attention is together decided on by bottom-up significance and top-down guidance.Notice that competition is first the thickest Yardstick under, and not noticed to cross target on carry out.

In sum, vision noticing mechanism is incorporated in flow production line and carries out target part identification and location, can make Industrial robot has greater flexibility in operation process.And development and application vision noticing mechanism will assist in and more enters One step ground improves the level of intelligence of industrial robot.

Summary of the invention

It is an object of the invention to overcome at present on industrial production line, manufactured parts needs by traditional artificial Identify and the deficiency of hand assembled, it is provided that a kind of versatility, robustness, concurrency, the more preferable view-based access control model of the suitability note The target part recognition methods of mechanism.

The purpose of the present invention is achieved through the following technical solutions:

A kind of target part recognition methods of view-based access control model attention mechanism, described target part recognition methods includes:

Step S1, vision noticing mechanism Model Selection, select the attention mechanism model of feature based and note based on space Meaning Mechanism Model, so as to utilize the central peripheral filter construction of biology on this basis, carry on multiple space scales Taking feature, described feature includes color, direction, brightness；

What step S2, feature combined significantly schemes to generate, and color, direction and brightness are combined into characteristic pattern, thus Obtain describing corresponding to the significance of color, direction, brightness, and the significance of these features is described through normalization meter Calculate and after linear combination, form notable figure；

Step S3, target manufactured parts recognition strategy, according to gather part image, generate notable figure, binary conversion treatment and Optimize, extraction marking area, the flow process of part zone extraction carry out Parts Recognition.

Further, the attention mechanism model of described feature based provides specific tasks with keyword form, will produce zero The color of part, direction, brightness as specimen, are made with the color of required manufactured parts, direction, the minimum deflection of brightness For significance, first define current manufactured parts by priori and store, then passing through the basic of study manufactured parts Feature, calculates the phase knowledge and magnanimity of basic feature and existing feature, thus detects maximally related manufactured parts, finally at target scene In find foreground, and mate with existing manufactured parts.

Further, described attention mechanism model based on space utilizes biological central peripheral filter construction, many Extract the color of manufactured parts, direction, brightness on individual space scale, then features described above is combined into characteristic pattern, from And obtain the significance description of the color corresponding to manufactured parts, direction, brightness, finally the significance of features described above is retouched State the notable figure through forming manufactured parts through normalization calculating and linear combination.

Further, described step S2, the notable figure generation of feature combination specifically include:

Step S21, generate notable figure based on color, brightness and direction character, input picture is carried out in different levels Nonuniform sampling, then passes through wave filter and extracts the color of different scale, brightness and direction character, the most again by each scale layer The characteristic pattern that eigentransformation is the multiple rank of same yardstick, then calculate its central peripheral after the recovery renormalization and obtain face Color, brightness and directional characteristic concern figure, final Fusion of Color, brightness and directional characteristic concern figure generate notable figure；

Step S22, carry out saliency value extraction based on rectangular histogram figure notable to comparison；

Step S23, algorithm based on regional correlation carry out saliency value extraction to notable figure.

Further, described step S21, based on color, brightness and direction character generate notable figure include:

S211, use gaussian pyramid model, different levels carry out nonuniform sampling, to a width input picture I (x, Y) following nonuniform sampling is carried out with gaussian pyramid G (x, y, σ):

$R(x,y,\mathrm{\σ})=I(x,y)\⊗G(x,y,\mathrm{\σ}),$

$G(x,y,\mathrm{\σ})=\frac{1}{2{\mathrm{\π\σ}}^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}^2}),$

In formula, σ is scale factor, namely the bandwidth of gaussian pyramid G (x, y, σ)；

S212, the extraction color of image, brightness, direction character, use central peripheral difference operator to carry out feature extraction, point Do not represent red, green, blue channel with r, g and b, then the brightness of image is expressed as

I (x)=[r (x)+g (x)+b (x)]/3

Component in original input image extraction four Color Channel redness, green, blueness, yellow: red R=r-(g+ B)/2, green G=g-(r+b)/2, blue B=b-(r+g)/2, yellow Y=(r+g)/2-r-g/2-b, direction character Use the component of four direction, wherein θ={ 0 °, 45 °, 90 °, 135 ° },

Each Color Channel of image is set up gaussian pyramid model, and obtains figure by central peripheral difference operator As the Feature Mapping figure on color characteristic, computational methods are as follows: utilize difference of Gaussian DOG (x, y) meter of center C and periphery S Nomogram is as I (x, Feature Saliency y)

$DOG(x,y)=\frac{1}{2{\mathrm{\π\σ}}_C^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}_C^2})-\frac{1}{2{\mathrm{\π\σ}}_S^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}^2})$

In formula, σ_cIt is the scale factor of center C, σ_sIt is the scale factor of periphery S, by higher level's image interpolation is amplified To side images, representing the calculating of central authorities C and periphery S difference with symbol Θ, central peripheral difference result of calculation is the pass of character pair Note figure: brightness figure I (c, s)=I (c) Θ I (s), color characteristic figure RG (c, s)=(R (c)-G (c)) Θ (G (s)- I (s)), BY (c, s)=(B (c)-Y (c)) Θ (Y (s)-B (s)), direction character figure O (c, s, θ)=O (c, θ) Θ O (s,θ)∣；

S213, feature concern figure is normalized and is generated final notable figure respectively, to through normalized N (I ((BY (c, s)) and N (O (c, s, θ)) uses for RG (c, s)), N for (c, s)), NComputing combines and obtains final notable figure, wherein, It is on different scale layer, the Feature Mapping figure of each feature to be carried out down-sampled, obtains the highest principal dimensions layer, then carry out Additive operation obtains the concern figure on color, brightness, direction characterWherein,

Brightness normalization characteristic figure

Color normalization characteristic figure

Direction normalization characteristic figure

Further, described step S22, carry out saliency value extraction based on rectangular histogram figure notable to comparison particularly as follows:

Each picture number in input picture is defined a saliency value, this saliency value by this as the color of number and its It represents as the color contrast of number, and in piece image, the saliency value of a pixel is defined as

$S\left(I_k\right)=\underset{\∀I_i\∈I}{\Σ}D(I_k,I_i)$

Above formula further expands becomes following form

S(I_k)=D (I_k,I₁)+D(I_k,I₂)+...+D(I_k,I_N)

In formula, N is the number of pixels in image, and the saliency value obtaining each color is as follows

$S\left(I_k\right)=S\left(c_i\right)=\underset{j=1}{\overset{n}{\Σ}}{f}_{j}D(c_1,c_j)$

In formula, c_iIt it is pixel I_kColor value, n is the pixel quantity of different colours, f_jBe image I (x, y) in color value be c_jNumber of pixels, by color quantizing with select to set up a succinct rectangular histogram by the way of frequency of occurrences color.

Further, described step S23, algorithm based on regional correlation notable figure is carried out saliency value extraction particularly as follows:

Divide the image into into regional first with image segmentation algorithm, then each region built color histogram, For a region, calculate its saliency value by calculating it with the color contrast in every other region in image

$S\left(r_k\right)=\underset{r_k\≠r_i}{\Σ}w\left(r_i\right)D_r(r_k,r_i)$

In formula, w (r_i) it is region r_iWeights calculate, D_r(r_k,r_i) it is the space length in two regions.Use region r_i In number of pixels be used as w (r_i), region r₁And r₂Between color distance be defined as

$D_r(r_1,r_2)=\underset{i=1}{\overset{n_1}{\Σ}}\underset{j=1}{\overset{n_2}{\Σ}}f(c_1,i)f(c_2,j)D(c_1,i,c_2,j)$

In formula, f (c_k, i) it is region c_kAll n_kThe frequency of i-th color, wherein k={1,2} in individual color.

Further, described step S3, target manufactured parts recognition strategy specifically include:

Step S31, collection part image also generate notable figure,

Industrial production line gathers a width part image, when extracting significant part zone from image, makes With based on rectangular histogram control methods, part being carried out significance detection, each pixel of piece image uses Color Statistical to determine Justice saliency value:

$S\left(I_k\right)=\underset{\∀I_i\∈I}{\Σ}D(I_k,I_i)$

In formula, D (I_k,I_i) it is I in L*a*b space_kAnd I_iBetween distance.For pixel I_i, its color is c_i, it is possible to To the saliency value of each color, above formula become

$S\left(I_k\right)=S\left(c_l\right)=\underset{j=1}{\overset{n}{\Σ}}{f}_i\×D(c_l,c_j)$

In formula, n is the color sum included in image, f_iFor color c in image I_jThe probability occurred；

Step S32, binary conversion treatment and optimization, extract for realizing object and noise separation and follow-up marking area, adopt Determine that threshold value carries out binaryzation with OTSU algorithm, use fixed threshold T ∈ [0,255] to carry out binaryzation；

Step S33, extraction marking area, use a boundary rectangle to take out the bianry image after optimizing in each region Take part marking area, by setting up convex shell and the minimum area area-encasing rectangle of the given 2D point set of rotational shell searching, i.e. For minimum enclosed rectangle, bianry image is divided into zones of different by these minimum enclosed rectangle, the corresponding different part in each region, Recording position and the size of each boundary rectangle, the same position at corresponding acquired original image adds an equal amount of square Battle array；

Step S34, part zone extract, and identify part by the geometric characteristic of part, i.e. use circular and side The circularity of shape shape, rectangular degree, area and girth feature identification go out different target part, and the eigenvalue of features described above is respectively It is defined as follows:

1. areaB is the matrix that part region bianry image is corresponding；

2. girth P, is around the length of region all pixels external boundary；

3. circularity E=4 π A/P²；

4. rectangular degree R=A/A_R, wherein A_RArea for minimum enclosed rectangle.

Step S35, Parts Recognition, identify whether there is target part on industrial production line.

Further, described step S35, Parts Recognition are the bianry image of each part of segmented extraction, define one Individual four-tuple aux_i=(A_i,P_i,E_i,R_i), wherein A_i, P_i, E_i, R_iRespectively the corresponding area of i-th part, girth, circularity, Rectangular degree, Parts Recognition algorithm is specific as follows:

S351, one target four-tuple target=(A of definition₀,P₀,E₀,R₀), for the characteristic vector of target part；

S352, aux to each part_i=(A_i,P_i,E_i,R_i), calculate the discrimination of itself and target part

${\mathrm{pro}}_i=\underset{j=1}{\overset{4}{\Σ}}{\mathrm{aux}}_i\left(j\right)-t\arg et\left(j\right)$

In formula, aux_iJ () represents j element in four-tuple；

S353, by all of discrimination pro_iCarry out ascending order arrangement, if the discrimination of minimum is more than a certain arithmetic number ε, Then think there is no target part on industrial production line, and provide information.Otherwise, by the district corresponding to the indexing of smallest region Territory is as target part region.

Further, in described step S32, binary conversion treatment and optimization, fixed threshold T is set between 80～100；

Described step S33, extraction marking area use rice profile minimum enclosed rectangle calculate grain type or utilize top The mode that some chain code combines with discrete Green's theory extracts the minimum enclosed rectangle of target image by method of principal axis and circumgyration.

The present invention has such advantages as relative to prior art and effect:

The target part recognition methods of a kind of view-based access control model attention mechanism disclosed by the invention, makes the vision of industrial robot System can efficiently identify the target part in work space, positions target part exactly, so that industrial robot exists Complete to have during manufactured parts assembling work higher autonomy, robustness, adaptability, go for industrial production line The upper detection of all kinds of parts, feeding, assemble, the occasion such as packaging.

Accompanying drawing explanation

Fig. 1 (a) is the target part identification process step schematic diagram of view-based access control model attention mechanism；

Fig. 1 (b) is vision noticing mechanism modular concept figure；

Fig. 2 (a) is vision attention example one；

Fig. 2 (b) is vision attention example two；

Fig. 2 (c) is vision attention example three；

Fig. 3 is that the feature integration noted is theoretical；

Fig. 4 is guiding search model framework；

Fig. 5 is Itti model framework；

Fig. 6 is task-driven attention model framework；

Fig. 7 is Sun-Fisher attention computation model framework；

Fig. 8 is target part identification process figure；

Fig. 9 is workbench part drawing；

Figure 10 is the notable figure generated；

Figure 11 is OTSU binary map；

Figure 12 is fixing threshold values binary map；

Figure 13 (a) is optimization process 1；

Figure 13 (b) is optimization process 2；

Figure 13 (c) is optimization process 3；

Figure 14 is binary map after optimization；

Figure 15 is bianry image marking area；

Figure 16 is the marking area gathering image；

Figure 17 is six sample parts；

Figure 18 degree of being to discriminate between scattergram.

Detailed description of the invention

For making the purpose of the present invention, technical scheme and advantage clearer, clear and definite, develop simultaneously embodiment pair referring to the drawings The present invention further describes.Should be appreciated that specific embodiment described herein, and need not only in order to explain the present invention In limiting the present invention.

Embodiment

Present embodiment discloses a kind of versatility, robustness, concurrency, the suitability more preferable view-based access control model attention mechanism Shown in target part recognition methods, concrete technical scheme such as Fig. 1 (a) and Fig. 1 (b).Described method includes three below step: S1, vision noticing mechanism Model Selection.Select the attention mechanism model of feature based and attention mechanism model based on space, with Just can utilize the central peripheral filter construction of biology on this basis, extract feature on multiple space scales, these are special Levy and include color, direction, brightness.What S2, feature combined significantly schemes to generate.Described color, brightness and direction character are combined Become characteristic pattern, thus obtained describing corresponding to the significance of color, direction, brightness, and notable by these features Property describe after normalization calculating and linear combination, form notable figure.S3, target manufactured parts recognition strategy.According to gathering zero The flow processs such as part image, the notable figure of generation, binary conversion treatment and optimization, extraction marking area, part zone extraction carry out part Identify.Separately below these three step is illustrated.

Step S1, vision noticing mechanism Model Selection；

Patent of the present invention selects the vision noticing mechanism model of feature based and vision noticing mechanism model based on space. The vision noticing mechanism model of feature based provides specific tasks with keyword form, first produces with priori definition is current Part also stores, and then passes through some basic features of study manufactured parts, calculates the phase of these features and existing feature Knowledge and magnanimity, thus detect maximally related manufactured parts, in target scene, finally find foreground, and with existing one A little manufactured partses mate.Vision noticing mechanism model based on space is first with biological central peripheral wave filter knot Structure, extracts the color of manufactured parts, direction, brightness on multiple space scales.Then these features are combined into spy Levy figure, thus the significance obtaining the features such as the color corresponding to manufactured parts, direction, brightness describes.These features notable Property the notable figure through forming manufactured parts through normalization calculating and linear combination is described.

S11, the attention mechanism Model Selection of feature based；

This model is as specimen using the color of manufactured parts, direction, brightness, with the color of required manufactured parts, Direction, brightness minimum deflection as significance.Thus can calculate significant spatial figure and identify dividing of manufactured parts Layer.It is to use top-down method that the attention mechanism model of described feature based describes the mode of required feature.Such as, extract Color, direction, brightness in one width manufactured parts image, obtain the notable figure under different characteristic.This model is with crucial font Formula provides specific tasks, first defines current manufactured parts by priori and stores, then passing through study manufactured parts Some basic features, calculate these features and the phase knowledge and magnanimity of existing feature, thus detect maximally related manufactured parts, finally In target scene, find foreground, and mate with more existing manufactured partses.

S12, attention mechanism Model Selection based on space；

So-called space refers to note the scene at object place or certain free surrounding space scope, and it is that a kind of common description is noted The mode of meaning power.This model thinks that attention is to carry out Selective attention with a certain particular range in free surrounding space when starting to act on , the visual stimulus in this spatial dimension can observed person be noticed, and the visual stimulus in other place can be automatically ignored. Famous scholar Itti etc. proposed the Itti visual attention model of classics in 1998 according to feature integration theory, as shown in Figure 7. Patent of the present invention based on this, first with biological central peripheral filter construction, is extracted raw on multiple space scales Producing the feature of part, these features include the color of manufactured parts, direction, brightness etc..Then these features are combined into spy Levy figure, thus the significance obtaining the features such as the color corresponding to manufactured parts, direction, brightness describes.Finally by these features Significance the notable figure through forming manufactured parts through normalization calculating and linear combination is described.

What step S2, feature combined significantly schemes to generate；

What feature combined significantly schemes to generate, and actually uses gaussian pyramid model to carry out nonuniform sampling, and extracts The features such as the color of manufactured parts image, brightness, direction, and normalized generate finally merged color, brightness and The concern figure of direction character these three feature generates notable figure.Notable figure be 1985 by relevant scholar in order to weigh piece image Significance and the concept of marked feature figure (be called for short notable figure) that puts forward.Notable figure is that a kind of phenogram is as vision attention district The two-dimensional distribution in territory, in notable figure, gray value shows that the most greatly the region significance of correspondence is the strongest, more can cause human vision The attention of system, the local maximum corresponding point of gray scale are referred to as the remarkable characteristic of image.If notable figure is compared to be one Map, then the place of significance maximum is equivalent to landform peak on map, and local feature region is equivalent to the small mountain of local, The region that physical features is the highest the most easily causes the concern of people.One width significantly schemes at least can to provide image where significance Relatively big and salient region scope has two information such as the widest.

S21, significantly scheme generate；

First, input picture is carried out nonuniform sampling in different levels, then, extracts different scale by wave filter Color, brightness and direction character.The most again by characteristic pattern that the eigentransformation in each scale layer is the multiple rank of same yardstick. Then, calculate its central peripheral after the recovery renormalization and obtain the concern figure of three features.Finally, these three feature is merged Pay close attention to figure and generate notable figure.Patent of the present invention uses classical Itti model generation significantly to scheme, wherein nonuniform sampling and extraction Two steps of feature are as follows:

S211, employing gaussian pyramid model, carry out nonuniform sampling in different levels.To a width input picture I (x, Y) following nonuniform sampling is carried out with gaussian pyramid G (x, y, σ):

$R(x,y,\mathrm{\σ})=I(x,y)\⊗G(x,y,\mathrm{\σ})$

$G(x,y,\mathrm{\σ})=\frac{1}{2{\mathrm{\π\σ}}^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}^2})$

In formula, σ is scale factor, namely the bandwidth of gaussian pyramid G (x, y, σ).

S212, the extraction color of image, brightness, direction character；

Because the significance of image is to embody the contrast in each characteristic dimension, it is possible to use central peripheral poor Operator carries out feature extraction.Represent red, green, blue channel with r, g and b respectively, then the brightness of image is expressed as

I (x)=[r (x)+g (x)+b (x)]/3

Component in original image extraction four Color Channel redness, green, blueness, yellow: red R=r-(g+b)/ 2, green G=g-(r+b)/2, blue B=b-(r+g)/2, yellow Y=(r+g)/2-r-g/2-b.Direction character is adopted With the component of four direction, wherein θ={ 0 °, 45 °, 90 °, 135 ° }.

Each Color Channel of image is set up gaussian pyramid model, and obtains figure by central peripheral difference operator As the Feature Mapping figure on color characteristic.Computational methods are as follows: utilize difference of Gaussian DOG (x, y) meter of center C and periphery S Nomogram is as I (x, Feature Saliency y)

$DOC(x,y)=\frac{1}{2{\mathrm{\π\σ}}_C^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}_C^2})-\frac{1}{2{\mathrm{\π\σ}}_S^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}^2})$

In formula, σ_cIt is the scale factor of center C, σ_sIt it is the scale factor of periphery S.By higher level's image interpolation is amplified To side images, represent the calculating of central authorities C and periphery S difference with symbol Θ.Central peripheral difference result of calculation is the pass of character pair Note figure: brightness figure I (c, s)=I (c) Θ I (s), color characteristic figure RG (c, s)=(R (c)-G (c)) Θ (G (s)- I (s)), BY (c, s)=(B (c)-Y (c)) Θ (Y (s)-B (s))；Direction character figure O (c, s, θ)=O (c, θ) Θ O (s,θ)∣。

Can be seen that from formula (8), central peripheral difference can reflect the height of image significance.After having obtained concern figure, There will be certain feature and there is the situation of many places contrast maximum, at this moment arise that substantial amounts of notable peak.If directly merged These features that there is a large amount of notable peak pay close attention to figure, will suppress to have notable peak other features less.So closing merging Before note figure generates notable figure, need it is normalized.

S213, feature concern figure is normalized and is generated final notable figure respectively.

For vision noticing mechanism, the attention of people can be affected mutually by too many high contrast region, and significance is on the contrary Reduce.Thus, it is desirable to be normalized the characteristic pattern that notable peak is few, it is a large amount of that Itti model uses normalization factor to weaken existence The characteristic pattern at notable peak.

To through normalized N, (((BY (c, s)) and N (O (c, s, θ)) uses for RG (c, s)), N for I (c, s)), NComputing is tied Conjunction obtains final notable figure.Wherein,It is on different scale layer, the Feature Mapping figure of each feature to be carried out down-sampled, To the highest principal dimensions layer, then carry out the concern figure that additive operation obtains on color, brightness, direction characterIts In,

Brightness normalization characteristic figure

Color normalization characteristic figure

Direction normalization characteristic figure

S22, notable figure is carried out saliency value carry based on rectangular histogram contrast (Histogram-based Contrast, HC) Take.

By the observation to biological vision attention mechanism, relevant scholar is found that organism is sensitive to visual signal contrast, Thus propose HC algorithm.This algorithm is a kind of notable figure extraction algorithm based on rectangular histogram contrast, every in input picture Individual all define a saliency value as number, this saliency value by this as the color of number and other represent as the color contrast of number. In piece image, the saliency value of a pixel is defined as

$S\left(I_k\right)=\underset{\∀I_i\∈I}{\Σ}D(I_k,I_i)$

In formula, D (I_k,I_t) it is as number I in L*a*b space_kAnd I_tSpace length.Above formula can also be expanded into as follows Form

S(I_k)=D (I_k,I₁)+D(I_k,I₂)+...+D(I_k,I_N)

In formula, N is the number of pixels in image.But being readily seen the most identical color value is to have phase With significance.Therefore identical color is calculated at one group, it is possible to the saliency value obtaining each color is as follows

$S\left(I_k\right)=S\left(c_i\right)=\underset{j=1}{\overset{n}{\Σ}}{f}_{j}D(c_1,c_j)$

In formula, c_iIt it is pixel I_kColor value, n is the pixel quantity of different colours, f_jBe image I (x, y) in color value be c_jNumber of pixels.In order to effectively calculate color contrast, set up by the way of color quantizing and selection frequency of occurrences color One succinct rectangular histogram.But quantization itself also can produce noise, some identical colors may produce after quantifying Different values.In order to reduce the noise jamming produced therefrom, need to use smoothing technique to redefine the aobvious of each color Work value.

S23, algorithm based on regional correlation (Region-based Contrast, RC)

In piece image, people often notice in image and produce the region of sharp contrast, except right with other regions Outside Bi, spatial relationship also can produce impact to human visual attention.And the contrast between other regions is the strongest, this region Significance the biggest.Scholar is had to propose RC algorithm.Spatial relationship is integrated into the calculating of region class by RC algorithm, uses each The sparse histograms control methods in region, divides the image into into regional first with image segmentation algorithm, then to each district Territory builds color histogram.For a region, by calculating it, the color contrast in every other region calculates with image Its saliency value

$S\left(r_k\right)=\underset{r_k\≠r_i}{\Σ}w\left(r_i\right)D_r(r_k,r_i)$

In formula, w (r_i) it is region r_iWeights calculate, D_r(r_k,r_i) it is the space length in two regions.In order to protrude from The contrast in other big regions, used here as region r_iIn number of pixels be used as w (r_i).Region r₁And r₂Between color away from From being defined as

$D_r(r_1,r_2)=\underset{i=1}{\overset{n_1}{\Σ}}\underset{j=1}{\overset{n_2}{\Σ}}f(c_1,i)f(c_2,j)D(c_1,i,c_2,j)$

In formula, f (c_k, i) it is region c_kAll n_kThe frequency of i-th color, wherein k={1,2} in individual color.This is the most just Be use color occur frequency as weights to the difference showing between primary color.In the rectangular histogram calculating each region Time, owing to each region contains only the sub-fraction color in entire image color histogram, so passing through conventional matrix Calculate and store histogrammic efficiency the lowest.Therefore, in order to preferably store and calculate, generally use sparse matrix generation Replace.

Step S3, target manufactured parts recognition strategy

The target manufactured parts recognition strategy that patent of the present invention uses as shown in Figure 8, namely on industrial production line Gather a width part image, according to gather part image, generate notable figure, binary conversion treatment and optimization, extraction marking area, zero Part region extraction, the process step of Parts Recognition are carried out.

S31, collection part image also generate notable figure.

Industrial production line gathers a width part image, as shown in Figure 9.Significant part is extracted from image During region, use based on rectangular histogram control methods, part is carried out significance detection.Vision is believed by the method based on biological vision Number observation that contrast is sensitive.To each pixel in image, its significance is by its color and the color of other pixels in image Contrast represents.Each pixel of piece image uses Color Statistical to define saliency value:

$S\left(I_k\right)=\underset{\∀I_i\∈I}{\Σ}D(I_k,I_i)$

In formula, D (I_k,I_i) it is I in L*a*b space_kAnd I_iBetween distance.For pixel I_i, its color is c_i, it is possible to To the saliency value of each color, above formula become

$S\left(I_k\right)=S\left(c_l\right)=\underset{j=1}{\overset{n}{\Σ}}{f}_i\×D(c_l,c_j)$

In formula, n is the color sum included in image, f_iFor color c in image I_jThe probability occurred.Because in industry The image gathered in production line is little, can process this in real time with existing PC.Generate significantly schemes such as Fig. 9 institute Show.

S32, binary conversion treatment and optimization.

Extract to realize object and noise separation and follow-up marking area, need Figure 11 is carried out binary conversion treatment. OTSU algorithm is generally used to determine that threshold value carries out binaryzation.OTSU algorithm is the simple high efficacious prescriptions of a kind of adaptive polo placement single threshold Method, image grey level histogram is divided into two parts of target and background with optimum thresholding, makes two parts inter-class variance take by the method Maximum, i.e. separation property are maximum.The method effect in target and background forms the image of obvious gray scale difference is notable, for Figure 10 The situation that middle part and background gray levels are more or less the same, separating effect is inconspicuous, easily causes Same Part fault-layer-phenomenon, such as figure Shown in 11 so that the minimum enclosed rectangle set up during extraction part can not comprise same part exactly.Special in the present invention In profit application, fixed threshold T ∈ [0,255] is used to carry out binaryzation in conjunction with practical situation.In order to reliably highlight marking area, Through test of many times, T is set between 0 to 255 change, sets the threshold to after contrast remarkable result compare between 80～100 Preferably, actual employing 80 is the most suitable, as shown in figure 12.

Meanwhile, in order to get rid of the noise impact on bianry image, in addition it is also necessary to be optimized.Specifically optimization process such as Figure 13 institute Show: if 1. the pixel value of current point is 0, then the next pixel of search, as shown in Figure 13 (a)；If the 2. pixel value of current point Be 1, and the upper right of this point, just go up, upper left, left front pixel value are all 0, then explanation runs into new target part.At image moment In Zhen, the number value on this position is that a upper non-zero number value adds 1, and the next pixel of search, as shown in Figure 13 (b)；If 3. The pixel value of current point is 1, and the upper right of this point, just go up, upper left, left front at least a pixel value be 1, the most in a matrix Placing the reference numeral that any one pixel value is 1 on this position, the next pixel of search, as shown in Figure 13 (c).So exist In matrix, multiple regions of existence difference numbering, different objects and noise in the most corresponding Figure 10, add up the individual of identical numbering Number, it is possible to obtain the area in this region, is set to region (this region corresponding noise section under normal circumstances) too small for area Background is removed, and effect is as shown in figure 14.

S33, extraction marking area.

Obtain the 3D position of part for convenience, the bianry image after optimizing is used an external square in each region Shape extraction part marking area.It is, for example possible to use rice profile minimum enclosed rectangle calculates grain type, or utilize summit chain code The mode combined with discrete Green's theory extracts the minimum enclosed rectangle of target image by method of principal axis and circumgyration.By building Vertical convex shell and rotational shell find the minimum area area-encasing rectangle of given 2D point set, are minimum enclosed rectangle.Such as Figure 15 Shown in.Bianry image is divided into zones of different by these boundary rectangles, the corresponding different part in each region.Record each external square The position of shape and size, the same position at corresponding acquired original image adds an equal amount of matrix, in order to prominent notable Region, can be set to different colours by boundary rectangle frame, as shown in figure 16.

S34, part zone extract.

Owing to major part part is to be made up of the simple geometric shape combination such as circular or square, therefore can several by part What shape facility identifies part, i.e. uses the feature identification such as circle and the circularity of square configuration, rectangular degree, area and girth Go out different target part.Make full use of shape and the geometric properties of part itself, it is possible to avoid sample training and feature The shortcomings such as the complexity that common-mode recognizer brings such as join.The eigenvalue that present patent application uses defines such as respectively Under:

1. areaB is the matrix that part region bianry image is corresponding；

2. girth P, is around the length of region all pixels external boundary；

3. circularity E=4 π A/P²；

4. rectangular degree R=A/A_R, wherein A_RArea for minimum enclosed rectangle.

S35, Parts Recognition.

For the bianry image of each part of segmented extraction in fig. 17, define four-tuple aux_i=(A_i,P_i, E_i,R_i), wherein A_i, P_i, E_i, R_iThe most corresponding area of i-th part, girth, circularity, rectangular degree.

Recognizer is as follows:

S351, one target four-tuple target=(A of definition₀,P₀,E₀,R₀), for target part characteristic vector.

S352, aux to each part_i=(A_i,P_i,E_i,R_i), calculate the discrimination of itself and target part

${\mathrm{pro}}_i=\underset{j=1}{\overset{4}{\Σ}}{\mathrm{aux}}_i\left(j\right)-t\arg et\left(j\right)$

In formula, aux_iJ () represents j element in four-tuple.

S353, by all of discrimination pro_iCarry out ascending order arrangement, if the discrimination of minimum is more than a certain arithmetic number ε, Then think there is no target part on industrial production line, and provide information.Otherwise, by the district corresponding to the indexing of smallest region Territory is as target part region.So can be carried out follow-up guiding industrial robot and mechanical hand carries out part assembling.

Carrying out component area indexing experiment, choosing 6 different parts is sample, as shown in figure 17.In order to observe same zero Part rotates the diversity of feature between different angle character situations of change and different part, calculates part 1～6 and zero respectively Part 1 rotates clockwise bianry image area, girth, circularity and the rectangular degree of 60 degree on industrial production line successively. For quantitative differences, the discrimination computing formula defined according to formula (11) calculates different part and Same Part rotates difference The discrimination of angle and target part 1.Result is as shown in table 1.

Table 1

Object	Area A	Girth P	Circularity E	Rectangular degree R	Discrimination
						Part 1 sample	26457	2192	0.0692	0.7800	--
Rotate 60 degree	26984	2077	0.0786	0.7867	0.0220
						Rotate 120 degree	24197	2554	0.0466	0.8213	0.0118
Rotation turnback	26147	2052	0.0780	0.7970	0.0209
						Rotate 240 degree	26846	2080	0.0780	0.7982	0.0195
Rotate 300 degree	24906	2501	0.0500	0.8068	0.0187
						Part 2 sample	22601	2848	0.0350	0.6860	0.3696
Part 3 sample	34736	2959	0.0499	0.6825	0.3138
						Part 4 sample	28166	2662	0.0499	0.6814	0.7394
Part 5 sample	6614	2802	0.0106	0.6976	1.3682
						Part 6 sample	4968	2802	0.0080	0.4772	1.6700

As it can be seen from table 1 the change of each eigenvalue is little after part 1 rotates different angles, calculate discrimination Value be both less than 0.1, part 2～6 is compared with part 1, the differentiation angle value calculated is both greater than 0.3.As shown in figure 18, Therefore can be 0.1 by the threshold definitions of discrimination.If being more than 0.1 by the differentiation angle value being calculated certain part, then recognize It not Same Part for this part and target part (nulling part 1), otherwise then think that this part is target part.

Above-described embodiment is the present invention preferably embodiment, but embodiments of the present invention are not by above-described embodiment Limit, the change made under other any spirit without departing from the present invention and principle, modify, substitute, combine, simplify, All should be the substitute mode of equivalence, within being included in protection scope of the present invention.

Claims (10)

1. the target part recognition methods of a view-based access control model attention mechanism, it is characterised in that described target part recognition methods Including:

Step S1, vision noticing mechanism Model Selection, select the attention mechanism model of feature based and attention machine based on space Simulation, so as to utilize the central peripheral filter construction of biology on this basis, extracts spy on multiple space scales Levying, described feature includes color, direction, brightness；

What step S2, feature combined significantly schemes to generate, and color, direction and brightness is combined into characteristic pattern, thus obtains Significance corresponding to color, direction, brightness describes, and the significance of these features is described calculate through normalization and Notable figure is formed after linear combination；

Step S3, target manufactured parts recognition strategy, according to gather part image, generate notable figure, binary conversion treatment and optimization, Extraction marking area, the flow process of part zone extraction carry out Parts Recognition.

The target part recognition methods of view-based access control model attention mechanism the most according to claim 1, it is characterised in that described base Attention mechanism model in feature provides specific tasks with keyword form, the color of manufactured parts, direction, brightness is made For specimen, with the color of required manufactured parts, direction, brightness minimum deflection as significance, first use priori Define current manufactured parts and store, then passing through the basic feature of study manufactured parts, calculating basic feature with existing The phase knowledge and magnanimity of feature, thus detect maximally related manufactured parts, finally find foreground in target scene, and with Existing manufactured parts mates.

The target part recognition methods of view-based access control model attention mechanism the most according to claim 1, it is characterised in that described base Attention mechanism model in space utilizes biological central peripheral filter construction, extracts manufactured parts on multiple space scales Color, direction, brightness, then features described above is combined into characteristic pattern, thus obtains the face corresponding to manufactured parts Color, direction, the significance of brightness describe, and finally describe the significance of features described above through through normalization calculating and line Property combination form the notable figure of manufactured parts.

The target part recognition methods of view-based access control model attention mechanism the most according to claim 1, it is characterised in that described step Rapid S2, the notable figure generation of feature combination specifically include:

Step S21, generate notable figure based on color, brightness and direction character, input picture is carried out in different levels non-all Even sampling, then passes through wave filter and extracts the color of different scale, brightness and direction character, the most again by the spy in each scale layer Levy the characteristic pattern being transformed to the multiple rank of same yardstick, then calculate its central peripheral after the recovery renormalization and obtain color, bright Degree and directional characteristic figure of paying close attention to, final Fusion of Color, brightness and directional characteristic concern figure generate notable figure；

Step S22, carry out saliency value extraction based on rectangular histogram figure notable to comparison；

Step S23, algorithm based on regional correlation carry out saliency value extraction to notable figure.

The target part recognition methods of view-based access control model attention mechanism the most according to claim 4, it is characterised in that described step Rapid S21, generate notable figure based on color, brightness and direction character and include:

S211, employing gaussian pyramid model, carry out nonuniform sampling in different levels, and to a width input picture I, (x y) uses Gaussian pyramid G (x, y, σ) carries out following nonuniform sampling:

$R(x,y,\mathrm{\σ})=I(x,y)\⊗G(x,y,\mathrm{\σ}),$

$G(x,y,\mathrm{\σ})=\frac{1}{2{\mathrm{\π\σ}}^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}^2}),$

In formula, σ is scale factor, namely the bandwidth of gaussian pyramid G (x, y, σ)；

S212, the extraction color of image, brightness, direction character, use central peripheral difference operator to carry out feature extraction, use respectively R, g and b represent red, green, blue channel, then the brightness of image is expressed as

I (x)=[r (x)+g (x)+b (x)]/3

Component in original input image extraction four Color Channel redness, green, blueness, yellow: red R=r-(g+b)/ 2, green G=g-(r+b)/2, blue B=b-(r+g)/2, yellow Y=(r+g)/2-r-g/2-b, direction character is adopted With the component of four direction, wherein θ={ 0 °, 45 °, 90 °, 135 ° },

Each Color Channel of image sets up gaussian pyramid model, and obtains image by central peripheral difference operator and exist Feature Mapping figure on color characteristic, computational methods are as follows: utilize difference of Gaussian DOG (x, y) the calculating figure of center C and periphery S As I (x, Feature Saliency y)

$DOG\left(x,y\right)=\frac{1}{2{\mathrm{\π\σ}}_C^2}\exp \left(-\frac{x^2+y^2}{2{\mathrm{\σ}}_C^2}\right)-\frac{1}{2{\mathrm{\π\σ}}_S^2}\exp \left(-\frac{x^2+y^2}{2{\mathrm{\σ}}^2}\right)$

In formula, σ_cIt is the scale factor of center C, σ_sIt is the scale factor of periphery S, obtains week by higher level's image interpolation is amplified Edge image, represents the calculating of central authorities C and periphery S difference with symbol Θ, and central peripheral difference result of calculation is the concern figure of character pair: Brightness figure I (c, s)=I (c) Θ I (s), color characteristic figure RG (c, s)=(R (c)-G (c)) Θ (G (s)-I (s)) , BY (c, s)=(B (c)-Y (c)) Θ (Y (s)-B (s)), direction character figure O (c, s, θ)=O (c, θ) Θ O (s, θ)；

S213, feature concern figure is normalized and is generated final notable figure respectively, to through normalized N (I (c, S)), ((BY (c, s)) and N (O (c, s, θ)) uses computing combination to obtain final notable figure to N, wherein, is for RG (c, s)), N The Feature Mapping figure of each feature is carried out down-sampled by different scale layer, obtains the highest principal dimensions layer, then add Method computing obtains the concern figure on color, brightness, direction characterWherein,

Brightness normalization characteristic figure

Color normalization characteristic figure

Direction normalization characteristic figure

The target part recognition methods of view-based access control model attention mechanism the most according to claim 4, it is characterised in that described step Rapid S22, carry out saliency value extraction based on rectangular histogram figure notable to comparison particularly as follows:

Each picture number in input picture is defined a saliency value, this saliency value by this as the color of number and other picture The color contrast of number represents, in piece image, the saliency value of a pixel is defined as

$S\left(I_k\right)=\underset{\∀I_i\∈I}{\Σ}D(I_k,I_i)$

Above formula further expands becomes following form

S(I_k)=D (I_k,I₁)+D(I_k,I₂)+...+D(I_k,I_N)

In formula, N is the number of pixels in image, and the saliency value obtaining each color is as follows

$S\left(I_k\right)=S\left(c_i\right)=\underset{j=1}{\overset{n}{\Σ}}{f}_{j}D(c_1,c_j)$

In formula, c_iIt it is pixel I_kColor value, n is the pixel quantity of different colours, f_jBe image I (x, y) in color value be c_j's Number of pixels, sets up a succinct rectangular histogram by the way of color quantizing and selection frequency of occurrences color.

The target part recognition methods of view-based access control model attention mechanism the most according to claim 4, it is characterised in that described step Rapid S23, algorithm based on regional correlation notable figure is carried out saliency value extraction particularly as follows:

Divide the image into into regional first with image segmentation algorithm, then each region is built color histogram, for One region, calculates its saliency value by calculating it with the color contrast in every other region in image

$S\left(r_k\right)=\underset{r_k\≠r_i}{\Σ}w\left(r_i\right)D_r(r_k,r_i)$

In formula, w (r_i) it is region r_iWeights calculate, D_r(r_k,r_i) it is the space length in two regions.Use region r_iIn Number of pixels is used as w (r_i), region r₁And r₂Between color distance be defined as

$D_r(r_1,r_2)=\underset{i=1}{\overset{n_1}{\Σ}}\underset{j=1}{\overset{n_2}{\Σ}}f(c_1,i)f(c_2,j)D(c_1,i,c_2,j)$

In formula, f (c_k, i) it is region c_kAll n_kThe frequency of i-th color, wherein k={1,2} in individual color.

The target part recognition methods of view-based access control model attention mechanism the most according to claim 1, it is characterised in that described step Rapid S3, target manufactured parts recognition strategy specifically include:

Step S31, collection part image also generate notable figure,

Industrial production line gathers a width part image, when extracting significant part zone from image, uses base In rectangular histogram control methods, part being carried out significance detection, each pixel of piece image uses Color Statistical to define aobvious Work value:

$S\left(I_k\right)=\underset{\∀I_i\∈I}{\Σ}D(I_k,I_i)$

In formula, D (I_k,I_i) it is I in L*a*b space_kAnd I_iBetween distance.For pixel I_i, its color is c_i, it is possible to obtain every Plant the saliency value of color, above formula become

$S\left(I_k\right)=S\left(c_l\right)=\underset{j=1}{\overset{n}{\Σ}}{f}_i\×D(c_l,c_j)$

In formula, n is the color sum included in image, f_iFor color c in image I_jThe probability occurred；

Step S32, binary conversion treatment and optimization, extract for realizing object and noise separation and follow-up marking area, use OTSU algorithm determines that threshold value carries out binaryzation, uses fixed threshold T ∈ [0,255] to carry out binaryzation；

Step S33, extraction marking area, use a boundary rectangle extraction zero in each region to the bianry image after optimizing Part marking area, by setting up convex shell and the minimum area area-encasing rectangle of the given 2D point set of rotational shell searching, is Little boundary rectangle, bianry image is divided into zones of different by these minimum enclosed rectangle, the corresponding different part in each region, record The position of each boundary rectangle and size, the same position at corresponding acquired original image adds an equal amount of matrix；

Step S34, part zone extract, and identify part by the geometric characteristic of part, i.e. use circular and square shape The circularity of shape, rectangular degree, area and girth feature identification go out different target part, and the eigenvalue of features described above defines respectively As follows:

1. areaB is the matrix that part region bianry image is corresponding；

2. girth P, is around the length of region all pixels external boundary；

3. circularity E=4 π A/P²；

4. rectangular degree R=A/A_R, wherein A_RArea for minimum enclosed rectangle.

Step S35, Parts Recognition, identify whether there is target part on industrial production line.

The target part recognition methods of view-based access control model attention mechanism the most according to claim 8, it is characterised in that described step Rapid S35, Parts Recognition are the bianry image of each part of segmented extraction, define four-tuple aux_i=(A_i,P_i,E_i, R_i), wherein A_i, P_i, E_i, R_iThe most corresponding area of i-th part, girth, circularity, rectangular degree, Parts Recognition algorithm is concrete As follows:

S351, one target four-tuple target=(A of definition₀,P₀,E₀,R₀), for the characteristic vector of target part；

S352, aux to each part_i=(A_i,P_i,E_i,R_i), calculate the discrimination of itself and target part

${\mathrm{pro}}_i=\underset{j=1}{\overset{4}{\mathrm{\Σ}}}{\mathrm{aux}}_i\left(j\right)-t\arg et\left(j\right)$

In formula, aux_iJ () represents j element in four-tuple；

S353, by all of discrimination pro_iCarry out ascending order arrangement, if the discrimination of minimum is more than a certain arithmetic number ε, then it is assumed that There is no target part on industrial production line, and provide information.Otherwise, using smallest region index corresponding to region as Target part region.

The target part recognition methods of view-based access control model attention mechanism the most according to claim 8, it is characterised in that

In described step S32, binary conversion treatment and optimization, fixed threshold T is set between 80～100；

Described step S33, extraction marking area use rice profile minimum enclosed rectangle calculate grain type or utilize summit chain The mode that code combines with discrete Green's theory extracts the minimum enclosed rectangle of target image by method of principal axis and circumgyration.