Underwater Image Based Fish Detection And Recognition Using Deep Learning Algorithm

The underwater object detection is important for monitoring the sustainability of the marine ecosystem.Nevertheless,the conventional monitoring system requires lots of resources because a marine ecosystem is difficult to access.Currently,the computer vision based underwater object detection can be applied to solve the problem.This thesis aims to detect the fish and recognize its type at the same time.Some challenges emerged in detecting and recognizing the fish type,such as poor image quality(unclear,blurry and small images resolution),the similar fish shape,the movement of fish in the water,and the limited fish datasets.We propose to use deep learning algorithms,like YOLO and Faster R-CNN,for detecting and recognizing the fish objects.And this is the first time YOLO version 3 algorithm applied for fish detection and recognition.Darknet53 and VGG16 are applied to YOLO version 3 and Faster R-CNN respectively as the feature extraction.Two datasets were used including fish4 knowledge and bubble vision.The fish4 knowledge dataset set is a public data set that has several neatly arranged images from each class,while the bubble vision dataset consists of videos that must be preprocessed before being used in the experiment.In order to get the quality image in the dataset,image preprocessing is applied using image enhancement methods including Image Filter(Homomorphic Filter and Gaussian Filter),Contrast adaptive histogram equalization(CLAHE),Multifision(CLAHE + Image filter),Dark Channel Priority.The experiment results showed the best result for fish detection and recognition obtained the performance of 99.16% using Faster R-CNN with multifision(CLAHE + Image filter)image enhancement on Bubble Vision datasets.While YOLO version 3 only performed 74% without image enhancement.