DE19718410C2 - Distribute real-time image data streams - Google Patents

Description

The result of quality control with image processing results often the problem that to cope with the necessary Computing power multiple processors are required because of the incoming data stream overloaded a single processor. At the same time, a system often consists of several side by side arranged signal or object sensors, such as cameras, Microphones, voltage scanners to the necessary resolution and Realize scanning width. The processing power of a Processor is usually not on the one to be delivered Data volume of a single camera matched.

So far - in the prior art - in image processing a structure according to 2a and 2b are used. For quality control systems. One or more sensor sources (here cameras) supply a processor (node) with data. In the case of very slow data streams, the data from several sensor sources (sensors) can also be completely processed by one processor. The computing power of the processor is designed so that it can process the data from all connected sensor sources. The billable amount of data of the processor must be greater than or equal to the amount of data supplied by the sensors. However, since today's cameras can deliver similar or larger amounts of data than the processor can process, it helps to insert a hardware-implemented preprocessing that reduces the amount of data to the computing power of the processor through suitable data loss. The permitted data loss in preprocessing is always dependent on the evaluation used, so that if the evaluation method changes, new hardware preprocessing must also be implemented. A software parameterization and a quick change of the processing algorithm is impossible due to the adapted required hardware preprocessing.

From Photonics Spectra, 3/97, page 41, is one with image data working multiprocessor system "Matrox Genesis" known. It is  scalable in its hardware, but on a bus connected together.

It is the task of the invention with a Processing system specifically to propose images that it allowed a processor system in a processing task optimally used with regard to the available computing power.

It is suggested that the data streams of all transducers merge and - adapted to the computing power of a each of several processors - this processor one to supply a certain amount of partial data for processing (claim 1 or 11).

The idea is to record the data of all sensors (sensors or Sources) in a picture distributor to a virtual one Unite total data stream and this collected data stream then flexibly on the connected processor nodes distributing extract. This gives each processor node exactly the amount of data he can still process.

According to the invention, this makes optimal use of the individual Processors possible, computing power no longer has to more or less exactly fit a transducer and vice versa; the computing power depends on the resolution of the transducer (Sensors), which the delivered (and to be processed) Specifies data performance, decouples. Unnecessary redundancy additional processor nodes and insufficient utilization a processor node can be avoided. The hardware will "scalable" and can consist of standard components such as Workstations or standard PCs. At a Expansion of the system based on a different algorithm and required higher computing power or more or sensors that are more willing to deliver data can use an additional one Processor nodes added and the distribution of data streams can be reconfigured without having to re-dimension the entire system to have to.  

A minimization of the processor nodes and a very good one This enables utilization of the computer capacities. at a switch to another processing algorithm only has to the distribution function in the image distributor can be changed. On the one hand, this allows an optimal degree of utilization multiple signal sources and also creates the opportunity a single sensor source the data on several Distribute processor nodes and the processor systems To make available without a data suppressive To have to do preprocessing.

The number m of outputs (output channels) and the amount of data each output channel in the image distributor is preferably programmable. Should a new distribution function for an existing system or such a scheme can be used and the computing power for this algorithm in a node is no longer sufficient an additional processor node included, the number of Outputs increased and the amount of data per output channel reduced.

The invention (s) are described below with reference to several Exemplary embodiments explained and supplemented.

Figure 1 is a generalized block diagram of an example of the invention.

Fig. 2a, Fig. 2b are generalized block diagrams according to the prior art.

Fig. 3 is a specific first block diagram of an image processing example of the invention with line scan cameras.

Fig. 4 is an internal block diagram of an example of an image distribution circuit.

Figure 5 is a second specific block diagram of an image processing example of the invention with area scan cameras.

The method will be explained and supplemented using the example of surface analysis. The problem was to carry out an endless inspection of a web material - not shown - with high resolution. This is illustrated in FIG. 3, which shows a special design of the scanning system aligned with images with union stage 20 a and extraction stage 20 b. Fig. 1 is.

Consider a web material that is 1200 mm wide and should be examined for defects of 40 µm in size. The max. Feed speed V _max is 600 mm / s. The scanning should take place with three line cameras 10 , 11 , 12 .

Assuming that the minimum error in each direction should be contained in at least two pixels of the camera the necessary resolution of the camera system 20 µm. For sampling of the web width you need 6000 pixels for one  Line frequency of 30 kHz (= 600 mm / sec divided by 20 µm). The The total data rate is 180 Mpixel / sec. As a camera system offers the use of three line cameras with 2048 pixels with two outputs and 30 Mpixels per output channel. Six channels of 30 Mpixel / s would therefore be the inputs to the image distributor to meet the above requirement.

To achieve the highest possible data bandwidth, own CCD sensors today may have more than one output channel (typically 1 to 16 channels), since the data rate per channel at max. 40 Mpixel / s lies. Then over each output channel certain part of the picture transmitted to the outside.

Multipentium boards 40 , 41 ,... Are used as processing units (processor nodes) in the example. , , 48 selected, in which each processor based on the processing algorithm a data rate of 20 Mpixel / s at your input e ₄₀ , e ₄₁ ,. , , e ₄₈ can process. For processing, 9 processors 40 to 48 are used, which are fed with parts of the virtual line. Each processor receives a partial line of 6000/9 pixels, i.e. approx. 667 pixels, which it can process. The results are exchanged between the processors via a network connection or forwarded to a master processor for control.

The image distributor 20 has gem. Fig. 4 two crossbar switch (crosspoint switch) 20 a and 20 b at the input and at the output. At the entrance of the data paths through the crosspoint switch 20 a on an internal storage structure of the image distributor to be mapped or pooled (conglomerates). Two memory banks are shown as memory 21 . A two-port memory (dual-port RAM), which can be read or written from two sides simultaneously, is available as the memory 21 .

A left address sequencer 22 a controls the crosspoint switch 20 a pixel by pixel and outputs the address for storage to the dual-port RAM 21 . At the output there is also a crosspoint switch 20 is b, the data read from the dual port RAM 21 pixel by pixel maps or from the internal memory structure to the output channels distributed (extracted). For this purpose, a second address sequencer 22 b is provided, which organizes reading and mapping of the output data independently of the first. A sequence controller 23 is used to synchronize the input and output data.

A virtual image formed by the controlled merging in the internal memory can be both a line and an area of the observed object. In one case, partial lines are cut out and made available to the processors 40 to 48 , in the other case certain partial areas can be passed on to the processors 40 to 48 for processing. This is preferred for tasks in the texture analysis of surfaces.

As an additional option, input data can be transferred to output channels a ₄₀ , a _41,. , ., a _{48 can be} distributed. In particular, overlapping regions can be created in this way, which can be evaluated and taken into account by several processors. Errors that normally lie on the boundary between two processing sections and are only partially recorded and evaluated by two different processors can be reliably detected by the overlapping processing regions and evaluated independently of one another.

According to another example. FIG. 5 shows the distribution of partial areas to a plurality of processors 40 to 43 . In a texture analysis Problem an area camera 19 detects the surface and directs the captured images with a pixel rate of 15 Mpixel / s to the image distributor 20 (z. B. constructed gem. Fig. 4).

Since texture analysis is very computationally expensive, a processor can only process an average data rate of 3.75 Mpixel / sec. Therefore, the image becomes partial areas of 64 × 64 pixels cut out and distributed to different processors. This can then make an evaluation in parallel on their subareas make and the results to a higher level  hand off.

The image distribution structure 20 can - starting from the general block diagram according to Fig. 1 - also extend to other fast signal sources for which the computing power of a processor is insufficient (e.g. acoustic signals, voltages, logic analysis, etc.).

The merging of the evaluation results from the individual Processors at the output of the image distributor are made in one "Master" to an overall result.

A multiple evaluation of the same data by one Parallel transmission to several processors is possible; it could use different detection and evaluation methods for the same surface area can be used.

Claims (12)

1. A method for distributing real-time image data streams from a plurality of signal sources ( 10 , 11 , 12 ,...; 19 ) using an image distributor ( 20 ) to a plurality of processor systems ( 40 ;...; 48 ) for processing
wherein the real-time image data streams in the image distribution under control of a first Adressensequencers load (22 a) into an intermediate memory (21), from which it as addressed by a second Adressensequencer (22 b) to the plurality of processor systems (40;...; 48 ) are issued;
and the control via the second address sequencer ( 22 b) depends on the respective processing power of the processor systems ( 40 ;...; 48 ). 2. The method according to claim 1, wherein the signal sources ( 10 ;...; 19 ) are image converters, in particular a two-dimensional area camera ( 19 ), for recording and outputting the image data. 3. The method according to claim 1, wherein the distribution of the output real-time image data streams via a respective individual coupling of a respective processor system ( 40 ;...; 48 ) to a respective separate computer connection (a ₄₀ ;...; A ₄₈ ) , 4. The method according to claim 1, wherein the output of the stored image data as a respective partial data stream comprises a respective partial data amount that is smaller than the amount of data temporarily stored in the buffer ( 21 ) of image data from the real-time image data streams of the plurality of signal sources ( 10 , 11 , 12 , .. ; 19 ). 5. The method of claim 1, wherein the plurality of processor systems ( 40 ; ... ; 48 ) are standard components, such as workstations or standard PCs. 6. The method according to claim 1, wherein when a processing algorithm is changed in one of the processor systems, the distribution of the data streams from the buffer store ( 21 ) is reconfigured or changed. 7. The method according to claim 1 or 3, wherein the number (m) of the outputs of the image distributor ( 20 a, 22 a, 21 , 20 b, 22 b) is variable, in particular programmable. 8. The method according to claim 1, wherein the real-time image data streams from the plurality of signal sources ( 10 , 11 , 12 ,...; 19 ) reach the buffer store ( 21 ) without loss of image data and the sum of the first data streams to the processor systems ( 40 ; ... ; 48 ) contain no less data than the sum of the data streams to the buffer ( 21 ). 9. The method according to claim 1, wherein in each of the processor systems ( 40 ;...; 48 ) image data directed there as a partial data stream are processed according to image processing methods and results of the processing of the processor systems as separate computer nodes via a network connection between the processor systems ( 40 ;...; 48 ) can be exchanged or forwarded to a master computer. 10. The method of claim 9, wherein the image processing methods are one or more of the following:
Object recognition, surface analysis, analysis of complex textures, person identification. 11. System for carrying out the method according to claim 1 for passing on blid data streams from signal sources ( 10 , 11 , 12 to 18 , 19 ) to processor systems ( 40 ;...; 48 ) which carry out processing, with a plurality of inputs for the image data streams which System has:

• a) an image data combiner ( 20 a, 22 a) with a first address sequencer ( 22 a);
• b) an image data distributor ( 22 b, 20 b) with a second address sequencer ( 22 b) and
• c) a memory ( 21 ) arranged between the image data combiner and the image data distributor for storing the image data streams supplied by the first address sequencer ( 22 a), the memory being controllable by the image data combiner ( 20 a, 22 a) and the memory being able to be fed by the same Image data distributor ( 22 b, 20 b) can be read out in a controlled manner;
• d) the two address generators being independent of one another and the reading being set as a function of a configured processing power of the processor systems ( 40 ; ... ; 48 ).

12. The system of claim 11, wherein a plurality of line or at least one area camera ( 19 ) serve as signal sources.