Event

 

 

Professor Qu Gang (Chair)

Professor Ang Li

Dr. Kleio Baxevani

 

Date/time: Monday, November 16 2026 at 10:00 AM 

 

Location: A.V. Williams (AVW) Room 2224

 

Title: An Automated Pipeline for Detection, Counting, and Length Estimation of Invasive Carp Using Recreational-Grade Side-Scan Sonar Imagery 

 

Abstract: Invasive carp pose significant challenges to aquatic ecosystem management, yet monitoring their populations remains difficult using conventional sampling methods. Although sonar offers an alternative for observing fish in turbid waters, research-grade sonar systems can be costly and difficult to deploy on a large scale. Recreational-grade side-scan sonar provides a more affordable alternative for fisheries monitoring. While previous studies have investigated individual tasks such as fish detection, counting, and length estimation, relatively limited attention has been given to integrating these capabilities into a single automated pipeline.

This thesis presents an automated pipeline for detecting, counting, and estimating the lengths of invasive carp using recreational-grade side-scan sonar imagery. Using a dataset collected by the United States Geological Survey (USGS), the proposed pipeline combines deep learning with traditional image processing methods. Full-resolution sonar images are divided into overlapping tiles and processed using a fine-tuned YOLOv12s-seg instance segmentation model to detect individual fish and generate segmentation masks. The predicted masks are subsequently refined using a Dense Conditional Random Field (Dense CRF), with parameters tuned for small fish targets and the intensity characteristics of sonar imagery. Fish lengths are then estimated using the minimum-area bounding rectangles of the refined masks. The number of fish is determined by counting the final bounding-box predictions produced by the detection model. These methods are integrated into a single pipeline that produces fish detections, segmentation masks, counts, and estimated lengths.

Experimental results demonstrate the performance of the proposed pipeline in fish detection, segmentation, counting, and length estimation. The model achieved a bounding-box mAP50 of 0.877. With tuned Dense CRF refinement, the pipeline achieved a mask precision of 0.773 and recall of 0.761. The refinement also reduced the mean absolute error (MAE) of fish length estimates from 76.95 mm to 48.01 mm, compared with reference lengths derived from ground-truth masks. Fish counting achieved an MAE of 4.12 fish per image tile.

Future work will focus on combining tile-level detections into full-resolution image results, removing duplicate detections, and incorporating geographic localization to support the mapping and monitoring of invasive carp populations.