Detection of Bluefin Tuna by Cascade Classifier and Deep Learning for Monitoring Fish Resources

Abstract

Detecting of specific species by image and video is a challenging task for cost-effectively fish monitoring systems. The current technologies to detect and count fishes for aquaculture are mainly based on sonar or manpower. The key problems to count fish in a fish cage using underwater optical cameras are uneven light distribution and fish overlapping. It also limits the application of traditional computer vision (CV) technologies. During recent years, the application of convolutional neural network (CNN) provides a deep-learning-based solution for fish monitoring. In this paper, an effective fish detection model based on region-based convolutional neural network (RCNN) is presented for counting Bluefin tuna in a fish cage, and a Haar feature-based cascade classifier model is provided for performance comparison. For model development, fish images captured by an underwater camera are extracted for training fish detection model. When we analyze a simple image in a fish cage as a primitive test, the RCNN model with 200 pieces of bluefin tuna image dataset in a fish cage has achieved the detection rate of91.5-and the accuracy of 92.4 %, while Haar cascade classifier presents the result of 67.0 and 53.8%, which means a significant number of tuna was reduced considering the number of tuna analyzed, respectively. Moreover, the training of RCNN by a RTX2080Ti GPU took only 52 seconds, about 1/35 of processing time of cascade classifier which trained by an Intel Core i5 CPU processor. Thus, the deep learning method with RCNN model performs an efficient effect on detecting underwater fish image and providing reliable analysis for aquaculture, which could help farmers count fishes more accurately with a larger dataset in the same time.