Deep Convolutional Neural Networks for Fish Weight Prediction from Images
Yunhan Yang, Bing Xue, Linley K. Jesson, Matthew J. Wylie, Mengjie Zhang, Maren Wellenreuther
Abstract
Fish weight is an important performance trait in aquaculture, conservation, fisheries science and management since weight relates to the growth of individual fish in a particular environment. A power regression model is commonly used to explain the relationship between fish weight and length. However, this requires costly measurements of fish length. The present study applies machine learning techniques to predict fish weight from fish images, bypassing the length measurement step. In this study, we validate the feasibility of predicting fish weight from images directly. We use a convolutional neural networks (CNNs) based approach to predict fish weight from images by building regression models. The deep CNNs architecture VGG-11, ResNet-18 and DenseNet-121 are chosen to train the models. The fish images have different scales (length-pixel ratio) without including a ruler as a reference. The trained regressors of these three architectures reach R<sup>2</sup> 0.94, 0.95 and 0.96 on the test set. Our results support the feasibility of fish weight prediction with the CNNs model from images directly. The fish images look similar to humans, but CNNs regressors can detect the different fish weights. The CNNs regressors also can detect the fish images with different length-pixel ratios.