Discrimination of Specific Introduced Species of Waterside Used AdaBoost

Recently, exotic breaking down the ecosystem of native species has occurred on a world scale. It is difficult for huge cost and people to hang when the exoticism breeds, and to exterminate it though capture by the hand work is a main current now as the measures. Therefore, it is necessary to discover before the exoticism breeds, and to capture it. Then, the research to aim to discover the exoticism in the pond and the lake where this problem appeared especially remarkably with the camera automatically in this research was done. This time, an original object detector was made by using Adaboost as an amount of the Haar-Like feature and a learning algorithm as an amount of the image feature used when the face was detected, and the discrimination of the black bass especially put in question because of the destruction of the ecosystem was tried.


1．Introduction
These days, the owner of the pet as for whom the heart is not desert a pet or it releases a black bass illegally for pleasure of fishing, The problem that a native species decreases sharply has occurred by taking food or eating a native species from the native species in which the living thing which must be present in the area, and which is not has the same food and ecology environment.In particular, at the waterside, such as a pond and a lake, such change is seen notably.Unless an alien species is exterminated, protection of a native species is in a hopeless state.As long as the exoticism is not exterminated, the protection of native species is hopeless.Many people have gone putting the time leave to remove the exoticism now.Cost and people become huge as the scale grows though it pulls out, it looks for the water of the pond and the lake, and there is an exterminating method, too.Especially, when the fertility is strong, and breeding starts from the black bass by one degree, exterminating all black basses becomes difficult.Then, the method of preventing breeding was devised by exterminating it as a method of defending the ecology only of native species at an early stage that the exoticism had invaded.The method is to find immediately when the camera is set up in the water of the pond and the lake where exoticisms of the black bass etc. have not existed yet, and the exoticism invades and to exterminate it.There are various one in the method of recognizing the image with the camera now.This time especially Make an original object identification machine by using Haar-Like amount of feature as amount of image feature and AdaBoost as learning algorithm that when the face is detected, it is used.And, the identification of the black bass especially put in question because of the destruction of the ecosystem was tried.

2．Principle 2.1 AdaBoost
AdaBoost is a learning algorithm for pattern classification.Since there is an advantage from which discrimination performance that mounting is easy and high is obtained, it is applied briskly in recent years.In study by AdaBoost, the function for the class that belongs to identify an unknown input pattern (unstudy sample) when the exercise of the pattern that belongs to two or more classes that want to identify it (study sample) is given is obtained.AdaBoost makes an discrimination machine different as the weight of the study sample is changed one after another, and gives a final identification function by the decision by majority with the weight of these two or more identification machines.An individual identification machine or less is called an weakness discrimination machine, and the one that they were combined is called a strong discrimination machine.AdaBoost is a kind of the Boosting algorithm that succeeds in spending by using the sample by updating the weight of the sample to "Adaptive".Figure2.1 shows the discrimination machine obtained by AdaBoost.

Haar-Like feature
The value of the amount of the feature is expressed by the difference between the brightness mean value in white square and the brightness mean value in the black square.A variety of Harr-Like features are generated with covering change of a square position and the size in the image of the study sample as shown in Figure2.2 before study beforehand.
For instance, when the sample of 20×20 pixels is given, it is tens of thousands of feature generable.At the study stage of the weakness discrimination machine, the one with the smallest discrimination error rate is selected from a lot of feature candidates.An effective feature to discrimination is one by one selected repeating this.

Attentinal-Cascade
When it scans in the image in the search window, things except the object' being found is far higher for the object to be found that the probabilities usually.Therefore, "Is the detection window of non-object refused to the squid?" becomes the key to speed-up.Then, accuracy and the speed  are improved by connecting two or more strong discrimination machines made by the explanation to here.The one that the discrimination machines were connected is called Attentional-Cascade as shown in Figure2.4.The one judged to be non-object with the first discrimination machine ends without doing processing any more when the detection window that existed in the image was installed.When one is judged from eyes as a target, it is passed to the second discrimination machine.Similarly, the one judged that it is the second discrimination machine and non-object is played, and other images are passed to the third discrimination machine.In a word, it is adjusted that the criteria of the object becomes severe by the exclusion of a doubtful one after another image in the discrimination machine of the upstream by a loose criteria, and going to the downstream.These discrimination machines can play non-object image at high speed because it usually uses feature patterns of a number a little like the upstream among 120,000 feature patterns.

Flow chart of study and discrimination
Figure2.5 shows the flow chart of study and discrimination.3．Experiment

Experiment method
In the experiment, the image was identified with a tool attached to OpenCV.The haartraining tool can originally make the feature file used by the object detection by the Haar-Like feature with the tool for the machine study of the image.A large amount of positive images can be generated with createsamples tool automatically.In this experiment, 20 images that the black bass is turning to the right are made as shown in Figure3.1.100 pieces are generated respectively with createsamples tool automatically, and 2000 positive images are prepared.When the discrimination machine is made, 5000 negative images are made and used from 500 images where the black bass is not reflected.
It takes a picture of the black bass that is actually swimming with the USB web camera and the discrimination situation is examined.It is made to enclose the area in the photographic image recognized the detection image and the same with a square frame and to display to it on the display, and it confirms it by watching.

4．Conclusion 4．Conclusion
The discrimination machine that detected black bass's image was made in AdaBoost generally used to detect the face of the person in this research.It identified it excluding the black bass though it was understood to detect a part of the black bass from the experiment correctly occasionally.
First of all, it is necessary to improve the discrimination accuracy as a problem in the future.And, whether only the black bass is detected while two or more fish exist or only a part of the black bass is enclosed with a square frame in a word is examined.As a result, it is thought that it is necessary to increase the positive image volume of data by extracting fish's characteristic from all angles if there is a problem in the detectivity degree.And, when the fish in water is actually identified in addition, it is necessary to consider the impurity of water etc.