Underwater species-constrained fish detection using multi-frame image information
Underwater fish detection system has many use cases such as fish biodiversity monitoring, aiding fish farming management, and providing data for marine resource management. Computer vision has proved to be a suitable tool for this fish detection task, as it is a low-cost, reliable, and most import...
Saved in:
Main Author: | |
---|---|
Format: | Final Year Project / Dissertation / Thesis |
Published: |
2023
|
Subjects: | |
Online Access: | http://eprints.utar.edu.my/6249/1/CEA_2023_LYJ.pdf http://eprints.utar.edu.my/6249/ |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Summary: | Underwater fish detection system has many use cases such as fish biodiversity monitoring, aiding fish farming management, and providing data for marine resource management. Computer vision has proved to be a suitable
tool for this fish detection task, as it is a low-cost, reliable, and most importantly,non-intrusive method for fish detection compared to trawling and other damaging methods. Detecting underwater objects introduces additional challenges, especially in unconstrained environments.Deep learning method has proved to be a powerful machine vision technique due to its deep hierarchical structures. YOLOv5 is used as an initial detector due to l) its K-means clustering to select anchor box size and 2) its PANet detection head.
Despite its strength, the result solely on YOLOv5 can still be improved, especially in decreasing the number of False Negative (FIN).
We observed a research gap to improve detection performance when there are domain differences between training and testing data. We propose a method that aims to fill that gap. The proposed method integrates an auxiliary system with the original YOLOv5, which is successful in decreasing the number of FN, albeit introducing some False Positives (FP). The overall F I score has We observed a research gap to improve detection performance when there are domain differences between training and testing data. We propose a
method that aims to fill that gap. The proposed method integrates an auxiliary system with the original YOLOv5, which is successful in decreasing the number of FN, albeit introducing some False Positives (FP). The overall F I score has improved by 5.28%. This auxiliary system provides information to select low- confidence bounding boxes produced by YOLOv5, and thus it produces additional candidates (bounding boxes) for reducing FN probability. The first step in the auxiliary system is the Trail Image Formulation module, which constructs trail images that are domain-agnostic. A trail image contains the
information of several image frames, which is derived from the concept of Motion History Images (MH]). Next, the detector of the auxiliary system is a modification and we name it YOLO-Ang. It takes in a trail image and produces bounding box candidates for each object in every frame. YOLO- Ang also produces angle information associated with the aforementioned bounding boxes. The output from YOLO-Ang is then processed using a Clustering-module and a simple Fusion module. To produce the final bounding boxes. In our extensive experiments, we compared three types of trail images (MHI), two types ofYOLO-Ang, and two types of Clustering modules. The best version ofthe above variants is able to achieve over a 5% Fl score improvement.
|
---|