Identify a Specified Fish Species by the Co-occurrence Matrix and AdaBoost - Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing - page 96

In recent decades, an adaptive boosting algorithm called AdaBoost[18, 19] is
invented. This boosting algorithm does not require any prior knowledge about the
performance of the weak learning algorithm. In this method, many weak classifier
consists a strong classifier with a learning process to find a suitable weightiness of
each weak classifier. The detection speed is very fast and the performance is good
for realtime application.
In this work, AdaBoost algorithm is adopt for fish area detection.

2.2 Mechine Learning Process

2.2.1 Sample Creation

AdaBoost is a excellent pattern detection algorithm, but in most case, it is used for
face detection. There is no research use AdaBoost algorithm to detect a fish, thus
we need to make a image database by ourself for study file creation.
In this work we collect 10 black bass images(Fig. 1) from internet or take a photo
of experiment fish tank. And the processing tool is with OpenCV library.
Ordinarily, the study process needs more than 5000 positive images and 3000
negative images. It is impossible for us to collect such amount images, here we use
the createsamples command included in OpenCV library to make 3000 forma-
tive sample images from these 10 images. Figure 1 show these image examples.

2.2.2 Training

In this work, machine learning process is also executed using an exist tool included
in OpenCV library called haartraining . By modifying the option parameters of
this command, finally, we got the study files in which recorded the weak classifier
and the weighting coefficient. We do not introduce the Training process in detail
because of the paper space. The reader can find the detail explanation on internet
easily.
The summary of the steps are as following

a. Target Image Collection:
In this work, only signal-object-taken images are selected, and the chosen image
background is also clear. we then put such images in to one folder on the compute
disk.
b. Negtive Image Selection:
Several background images without the detection target include are collected.
The size of the negative image larger than the positive image size so that each
teaching image have a different background.
c. Postive Image Creation
Applying the createsamples command with specify the target image and the
negative images that you prepared in a. and b., the positive images are created
automatically.
d. Postive Image Data Confirmation