Article Preview
Top1. Introduction
Subordinate level classifications are currently drawing a lot of attention from research community (Biederman, Subramaniam, Bar et al., 1999; Farrell, Oza, Zhang et al., 2011; Yao, Bradski, & Fei-Fei, 2012). The subordinate level categorization requires distinguishing among similar kinds of objects, such as computer tables and dining tables, rather than a basic level categorization, such as distinguishing between a table and a chair. These types of standard categorization is also known as “Fine-grained-Categorization” (Yao, Khosla, & Fei, 2011; Weinshall & Hillel, 2007; Yao, Bradski, & Fei-Fei, 2012) . Generic object classification includes the daunting task of differentiating amongst objects that only have subtle differences (Csurka, Dance, Fan et al., 2004).
It is predictable that traditional classification approaches cannot work well for Fine-grained-Categorization problems, because categorization and fine-grained categorization require different levels of fineness of features . For example, in the visual vocabulary based approach, the image patches are encoded using clustering which results in the loss of fine details, which are important for Subordinate level classification (Csurka, Dance, Fan et al., 2004; Lazebnik, Schmid, & Ponce, 2006; Rosch, Mervis, Gray et al., 2004; Wang, Yang, Yu, & Lv, 2010; Yao, Bradski, & Fei-Fei, 2012).
Species classification is one of the toughest sub-ordinate level classification problem. In nature there are different species for the same class of organism. There are numerous species in the class 'Fish'. Species are grouped under a 'Family' which are based on similar characteristics. Species from different families may significantly differ from each other whereas the species from same family can have many similarities; as a result, sometimes it is difficult to discriminate among them, as shown in Figure 1 and Figure 2. Some recent work in the area of subordinate level categorization deals with the classification of different datasets such as flowers (Zisserman & Nilsback, 2006), larvae (Martinez-Munoz et al., 2009), bird species (Martinez-Munoz et al., 2009; Yao, Bradski, & Fei-Fei, 2012) and leaf nodes of image net (Deng, Dong, Socher et al.). Through our exploration of the research in this area, we have not found any well known work in the area of image-based classifications of fish species.
The classification of fish species is a challenging task because the images of same fish species can show significant differences in the fish's attributes when taken in different conditions. For example; the colours of a fish in an underwater picture can look completely different when compared to a picture taken in the open environment. In Figure 2 there are two images of 'White Spotted Surgeon Fish' (Refer to Images 1and 2). These pictures are taken in an open environment, and under water conditions respectively. A large colour difference can easily be noticed in the two pictures. Additionally, we can observe a huge difference in the attributes in the images of same species, even when the photos are taken in same environment. The same fact can easily be observed in Images 3 and 4 in Figure 2, Both of the images are of 'Doctor Surgeon Fish', and are taken in underwater conditions; however, the difference in the attributes of the fish is huge. Usually such kinds of diversity are not observed in other species' datasets such as the 'Bird datasets of Caltech' (Wah, Branson, Welinder et al., 2011). The classification can be more complicated when images have different background as highlighted in the images of Figure 2.