Article Preview
Top1. Introduction
Automated underwater image classification has become most crucial challenge in the field of computer vision due to its image turbidity. Classifying images comes under computer vision based interdisciplinary field of artificial intelligence. Underwater turbid images have low-level features of image primitives. In case of ordinary images, classification algorithms work well and identify with good accuracy but for classifying turbid underwater images, system must perform higher and complex calculations. Underwater image classification system needs to deal with supplementary problems originated by the image degradation due to the light transmission in water. Underwater images endure from various challenging degradation effects like, absorption, scattering, marine snow or vignette which heavily influence visual perception (Lu et al., 2018; Li et al., 2017).
Absorption effect is the reason for the bluish color due to least attenuation in the medium of underwater images. Scattering effect modifies the direction of the light so that object will be blurred. Effect of vignette reduces the intensity of the image corner. Preprocessing is required to restore actual colors and enhance the image for next process. Image restoration and image enhancement techniques are used to resolve underwater image degradation problem. Applications of underwater images are 3D reconstruction of scenes, coral image classification. In order to process the underwater images with its applications, appropriate image processing algorithms are required to enhance the quality of underwater images. The main challenge of underwater image is not to have reference images to compare due to its turbidity (Lu et al., 2017; Salman et al., 2016).
Several methods for image classification was proposed and compared with other methods. But base idea for image classification is to extract information from the image by labelling the pixels of the image to dissimilar classes. Supervised classification and unsupervised classification are the two ways to classify the images. Unsupervised learning algorithms are working based on pixels of the image that are clustered into groups without interference of the analyst. The information of the clustered pixels is retrieved from the images and very less labelled data are available in reality, so unsupervised learning plays crucial role for image classification. Supervised learning technique uses labelled data for analysis, training the classifier model and feature extraction. Initially model can be trained using training dataset then it will classify the new images based on its features according to the information which learned in training phase (Jin & Liang, 2017; Zhu et al., 2017).
Present days, deep learning algorithms are very popular in the field of computer vision due to its tremendous computing capability to process the large dataset by learning the things deeper in order to get successful result. One of the application areas of machine learning is computer vision, which deals with image classification, object recognition, etc.., In general, Convolutional Neural Network (CNN) is being used in deep learning and it is suitable machine learning algorithm for image classification. CNN has the capacity to analyze the visual images exactly, since it applies feed-forward artificial neural network techniques (Chan et al., 2017; Wu et al., 2018).
Many conventional machine learning tools like Support Vector Machine (SVM), Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), have a tendency to saturate rapidly as the quantity of the training set grows drastically. Deep Learning targets to learn multiple levels of data representation, from low-level to high-level, to make sense of images. Deep convolutional neural networks, scalable computational resource with huge amount of training dataset (ImageNet) achieved tremendous accuracy in image classification (Krizhevsky et al., 2012).