Using machine vision to estimate fish length from images using regional convolutional neural networks

An image can encode date, time, location and camera information as metadata and implicitly encodes species information and data on human activity, for example the size distribution of fish removals. Accurate length estimates can be made from images using a fiducial marker; however, their manual extraction is time‐consuming and estimates are inaccurate without control over the imaging system. This article presents a methodology which uses machine vision to estimate the total length (TL) of a fusiform fish (European sea bass). Three regional convolutional neural networks (R‐CNN) were trained from public images. Images of European sea bass were captured with a fiducial marker with three non‐specialist cameras. Images were undistorted using the intrinsic lens properties calculated for the camera in OpenCV; then TL was estimated using machine vision (MV) to detect both marker and subject. MV performance was evaluated for the three R‐CNNs under downsampling and rotation of the captured images. Each R‐CNN accurately predicted the location of fish in test images (mean intersection over union, 93%) and estimates of TL were accurate, with percent mean bias error (%MBE [95% CIs]) = 2.2% [2.0, 2.4]). Detections were robust to horizontal flipping and downsampling. TL estimates at absolute image rotations >20° became increasingly inaccurate but %MBE [95% CIs] was reduced to −0.1% [−0.2, 0.1] using machine learning to remove outliers and model bias. Machine vision can classify and derive measurements of species from images without specialist equipment. It is anticipated that ecological researchers and managers will make increasing use of MV where image data are collected (e.g. in remote electronic monitoring, virtual observations, wildlife surveys and morphometrics) and MV will be of particular utility where large volumes of image data are gathered.