BIOMASS ESTIMATION OF FISH USING DEEP NETWORKS AND STEREO VISION

Abstract

Mass of an object is essential information for many an industrial application. The aquaculture industry estimates the fish's weight by scaling a sample of fish out of growing tanks. This process harms fish and reveals an inaccurate result. The research found there is a high correlation between fish weight and its size in the image. This study aims to use a convolution neural network (CNN) for estimating fish weight.

Firstly, according to our research, VGG19 or similar models were not tested to solve this problem before. Therefore, with known distance we tested CNN models VGG19 then compare its result with semantic segmentation models such as FC-Densenet, where another study applied a semantic segmentation technique on a smaller problem. To do this experiment, we used a fish dataset included 1275 images of harvest Salamon fish and their mass. The VGG19-R archived the lowest mean absolute percent error (MAPE), MAPE = 2.4\%, and the FC-Densenet-R revealed MAPE = 6.49\%.

To stimulate fish in a tank, we took a picture of a Lego block with a stereo vision camera in different positions to the camera. Then, we used the stereo data [right, left, depth map] as input to the VGG19-R model to estimate the area of the object. The model achieves MAPE= 2.37\% for the testing dataset. The result shows that the stereo vision camera could help to measure objects at different depths like fish inside the tank, where the depth map information works as a re-scaling factor to object area in the other inputs.