ABSTRACT
Fish species classification in underwater images is an emerging research area for scientists and research scholars in the field of image processing. But the phenomenon of light scattering and absorption in ocean water leads to hazy, dull and low contrasted images making fish classification a tedious and tough task. Moreover, the scarcity of available data set makes it further strenuous job to train a neural network from scratch. To overcome the above issues the present chapter proposes two automatic methods for fish categorization in underwater images. Initially, the first proposed FishNet method uses only last few layers of pre-trained AlexNet for classification of fish images. The FishNet has been trained on a limited data set of fish images without any data augmentation. The second method utilizes the architecture of ResNet50 to extract features from it and then feed them to Support Vector Machine classifier for fish species classification. Both the proposed methods have been tested on large as well as small data sets. Large and small data set comprises 27,370 (Fish4Knowledge) and 600 (QUT) fish images, respectively, having low resolutions and illuminations. Result analysis indicates that both the proposed methods give good accuracy of 98.67% and 95.90%, respectively, for large data set and 89.17% and 77.86%, respectively, for smaller data set.
