Underwater Fish Weight Estimation Using AI Length Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing manual methods for estimating fish average weight in aquaculture are invasive, laborious, inaccurate, and time-consuming, often failing to represent the entire population, leading to unreliable biomass data and inefficient resource management.
Innovation Solution
An automated system using underwater cameras and machine-learning models to detect and measure fish size, refining estimates with AI-based length calculations to determine accurate average fish weight, eliminating the need for costly stereovision cameras.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual sampling methods are used to estimate fish average weight, then the measurement process is simple and equipment-free, but the accuracy is poor (15-25% inaccuracy) and the sampling is invasive causing stress and physical damage to fish
Solution Approach 1:
The patent replaces manual mechanical sampling with an automated computer vision system using underwater cameras and machine learning models. The system captures images of fish in their natural environment, uses detection models to identify fish and generate bounding boxes, applies geometric corrections for camera perspective, and estimates lengths and weights non-invasively, eliminating physical contact and stress on fish while significantly improving measurement accuracy
Solution Approach 2:
The patent introduces an intermediary computational layer between image capture and weight estimation. The system uses detection models to generate virtual box-regions, applies geometric correction algorithms to account for camera distance and angle, and employs length estimation models to convert corrected box dimensions into accurate fish length measurements, which are then transformed into weight estimates through analytical models
2Measurement precision
If manual sampling is performed frequently to improve accuracy, then more data is collected, but the process becomes extremely time-consuming and laborious
Solution Approach 1:
The patent enables continuous monitoring by deploying underwater cameras that continuously capture video frames of fish in the aquaculture environment. The automated processing pipeline continuously detects fish, corrects geometric distortions, estimates lengths, and calculates average weights in real-time, allowing frequent or continuous measurements without additional manual labor or time investment
Solution Approach 2:
The system performs self-service by automatically processing captured images through the entire measurement pipeline without human intervention. The detection models autonomously identify fish and generate bounding boxes, the geometric correction algorithm automatically adjusts for camera perspective, the length estimation model computes fish lengths, and the system calculates average weights, eliminating the need for manual sampling operations
3Productivity
If a small sample size is used to reduce effort, then the process is faster, but the estimation becomes unreliable as it does not cover even 1% of the population
Solution Approach 1:
The patent applies partial action by processing only the fish visible in captured video frames rather than attempting to physically sample the entire population. The system captures images containing multiple fish, detects and measures all visible individuals in each frame, and aggregates data across numerous frames to achieve comprehensive population coverage without the laborious process of physically handling each fish
4Productivity
If the size of virtual box-region is used as initial proxy for fish size, then the measurement is rapid and automated, but the estimate is inaccurate due to camera distance and configuration variations
Solution Approach 1:
The patent changes the parameters used for length estimation by transitioning from raw bounding box dimensions to geometrically corrected dimensions. The system calculates correction factors based on camera intrinsic parameters, focal length, distance to the fish, and angle of view, applying these transformations to the virtual box-region dimensions to compensate for perspective distortion and produce accurate length measurements that maintain automation speed
Data Source
AI summary
The present disclosure relates to an automated method, computer program, and system for estimating the average fish weight in an aquaculture environment. An underwater camera in the aquaculture environment captures video frames from the environment. These frames are inputted into a machine learning detection model that has been trained to detect fish. The machine-learning vision detection model identifies a virtual box-region around detected fish and outputs size and location for the virtual-box regions, as well pixel data related to the virtual-box region and the frame as a whole. This size, location, and pixel data is processed and then inputted into an AI-based length estimation model, which computes a fish length for each virtual-box region inputted into the model. The system then calculates the average fish weight from the fish length using an analytical model.