Deep Learning for Fish Recognition & Measurementfeatured

Sponsored by Missouri Department of Conservation (MDC), we are working on fish recognition and measurement based on computer vision.

The problem: MDC fisheries biologists measure fish in Missouri waterways to get an understanding of the fish species, the waterway and the health of both. This can be a time-consuming task that involves net catch-and-release of multiple fish and the risk to fish due to the length of time out of water.

“The process to capture, store and document the fish takes a lot of time,” Sartwell said.

The solution: Doctoral students in Shang’s group are developing an app that will use a mobile device’s camera function to automatically capture fish measurements and determine the type of fish using deep learning, in particular convolutional deep neural networks, a bleeding edge machine learning technique for image and video recognition, which is also a major component in the success of Alpha Go, Google’s champion-level software for the board game “Go.” In the event the fish cannot be kept still long enough for a still photo, the group plans to develop video capturing capabilities that will use the longest length captured in a video.

Source: http://engineering.missouri.edu/2016/03/mdc-partnership/

Field work in a boat on Missouri River

People used to measure the length of fish by ruler, as shown in the picture.

We implemented AI to localize the mouth (red mark), eye (blue mark) and tail (green mark). With QR codes printed in known sizes as reference object, the AI can measure the length of fish automatically without a ruler.

Thanks Lin for her goldfish and photography.

The following video was recorded by myself in a local Hong Kong Market.