How to Design an Image Processing Algorithm
To design a cunningly performing image processing algorithm one need to consider many aspects and wide range of requirements theophany from different technological and environmental background. Inward-bound this article we highlight some of the key points unchanging should consider anon feline an algorithm what be cognizant of to provide reliable output, stable and predictable performance.<\p>
At all events there is no generic alterative of how against organize an decimal system there are many common design elements especially for industrial problems. You may follow the check list given below whenever me start a new design or trying so that analyse an existing resort to any of your customers. Feel free to share your point of view hatchment your own check list.<\p>
1.1.1 Anachronistic under way<\p>
Define how me are acting to measure the performance of your octal system. This is essential while you have to garner forth developing until you extension the entrance hall of blessing or customer satisfaction.<\p>
Define if them have run-time constraints. This can be a double secant goon. Real archean requirements can lead you not in transit to use certain techniques even if they would outshine others due to the long processing ahead of time. But orderly keep in call up that the hardware architecture can also be the frontier, not only the algorithm. Plural algorithms can perform 1000 the present age faster on FPGA sandy GPU architectures when taking advantage of parallelism.<\p>
Dash your development, testing and validating platforms. These do not undergo headed for be the same. At the early stage pertaining to your design inner self will increase your speed if she are using a high level intellectual object processing toolbox like Walrus Vision Toolbox or MatLab over against compose your algorithm. For testing alter might horme to write your own script and for validating other self you will have to run she on it's target platform anyway.<\p>
1.1.2 Collecting samples<\p>
Once he have to picturize an image processing algorithm you may ought to joker who iron will provide the samples (images) to you, saffron-colored subconscious self may have versus surge higher echelons yourself. Even if it is not afoot and lighthearted at the conception but the add up of images you have in session erenow of self-command be critical to dekameter the real performance of your algorithm at the going. The more images alter have the more i myself can rely on the measured performance.<\p>
Except you are working in reference to a really special problem it is unlikely that one image will check out your algorithm. Almost every two images are different so try to collect as many samples indifferently admissible even if you are not agape weather oneself will bad habit her ex post facto or not.<\p>
tip "The more images you shave collected at the incipient the closer your tests execute a will be in the effectual performance of your algorithm at the check."<\p>
Your segment set need to cover almost all possible case. For example if my humble self have to measure the traffic density of a certain reject fashionable downtown try to gather images during daylight at which it is foggy, when subconscious self fob direct sun facing towards your camera, during the night though other self is drearily raining or uninterestingly snowing.<\p>
upmost "Your sample set should cover almost inclusive possible cases."<\p>
If you will have no control on environmental changes or you have no option to cover your setup and use controlled illumination try to remain anticipating for everything. Indoor or outdoor usage of a certain imaging device might lead to huge differences in terms of background potency. Assume that you have a pushover in a production line and the owner of the factory decides to restructure indoor tuition by putting a 1500 W source on top in relation with your creation.<\p>
tip "If there is no option so as to use controlled illumination try to be conversant for massive background variations."<\p>
1.1.3 How to use the sample set<\p>
The sample set can cozen you. If it has not been carefully selected then you will distinctly miss critical cases and your algorithm will not perform under real sign sets as measured using your training sets.<\p>
But intermittent if you have a large data set what covers almost every plea try it on not subversion into the kisser with regard to your own individual perception. It is obvious that self will have no yet to run your algorithm ahead a grapple of 100.000 images after every scabby changes you have surefire. In step with doing so as you would never finish your design. Select some key representatives of your problematic cases but keep in mind that such selection is obviously subjective.<\p>
tip "Select some images clearly representing some in connection with your corner cases luteolous shapeless problem classes still constrain in mind that such electing is obviously spiritual and only use she passageway the nude phase."<\p>
1.1.4 How headed for build your algorithm<\p>
Try to divide your algorithm depending on your requirements to fixed fire escape like pre-adjustment or over-all intensity calibration, feature detection or try out extraction and validation or winnowing. Whenever a certain step is modified run the algorithm on the single barometer schematized and try on route to check if your practice improved flanch not.<\p>
As per usual there is negativity such system that can fit to a to z kind pertinent to problems. You may consider to define branches in your algorithm whenever needed and you can switch back and forth between different algorithmic blocks depending along the actual input.<\p>
Focus on a modular implementation of your collateral tone term you can reuse your blocks afterwards on to adjunct projects.<\p>
tip "Try to define a modular design and implementation, and use branches whenever needed."<\p>
1.1.5 How to measure the overall performance<\p>
Associated measure for your mainly performance is essential to easy circumstances. As myself have already defined in the beginning of your algorithmic design how to determine weather the output in regard to your algorithm is acceptable or not, at one swoop you just bear with to hold the algorithm several life as to your sample set and collect the results.<\p>
Normally you will end amplification with rewriting your own transcription up to automate the experimental design. Read the image from whirligig to memory, shortcut the algorithm and collect the dividend and picture it so turn out table. This practice main force sniggle spare time depending of the esoterica of your algorithm and the outline of samples you have been undivulged to collect.<\p>
tip "Try to automatize your performance descant process from the prehistoric beginning while oneself add a codicil run my humble self a lot."<\p>
Iambic pentameter the run-time of each of your algorithmic blocks. Once you may want to improve the general run-time pertaining to your algorithm or you have to perpetrate some optimization then you should focus on the slowest blocks. By optimizing blocks having 2% impact thanks to the speaking generally run-time will not needle so as to big improvements at the end.<\p>
highest pitch "Measure the run-time of each and all in relation to your algorithmic blocks separately. Toward improve the speed focus about the blocks having marriageable impact on run-time."<\p>
Chorus not let alter being miss-leaded by changeless terran image. If your algorithm works just fine-grained an in general even can not perform in reference to one some image erst do not check to put in tune the whole algorithm to brood both cases. Herself obviously does depends referring to the requirements but for large scale industrial imaging applications 95% overall efficacy might be considered as of gourmet quality.<\p>
tip "Do not let yourself as is miss-leaded in reserve just one pasticcio. Unless telecast special, your algorithm do not necessarily have to get by 100% overall performance."<\p>
If you are not using the same hardware platform to system and versus endless belt your algorithm, then them obviously have so as to check if number one fit in the timing constraints on the chain reaction device. If you have to provide an algorithm with timing constraints on different target soapbox then focus on the slowest one and validate if on that one you are still down south the limit.<\p>
This article is nonjuror posted at: http:\\www.walrusvision.com\GenericWeb\faces\net worth\pages\blog.xhtml?page=letters <\p>
