An Efficient Algorithm for Prawn Detection and Length Identification

Abstract

Prawn fishery has been gaining vast popularity in the aquaculture industry. But farmers fail to know health status of the prawn. Length and weight are parameters for assessing the health of the prawn. It is usually easier to measure the length of the specimen than the weight, and weight can be predicted using the length–weight relationship. In this application, Faster Region-Based Convolutional Neural Network (Faster RCNN) algorithm is used for the detection of the prawn and to draw a bounding surrounding the specimen. Faster RCNN returns coordinates in the form of [ymin, xmin, ymax, xmax] which can be used to localize prawn in the image. Pixel length is achieved using the above coordinates, and pixel per metric is used to derive length in centimeters from pixel length. The weight is achieved by applying length–weight relation. This weight is stored in the database. A graph with weight analysis is returned. In this work, we considered 100 images as test set and got an accuracy of 95% and ∓2 cms approximation in length.