A rapid, low-cost deep learning system to classify squid species and evaluate freshness based on digital images

Abstract

We developed and evaluated a rapid, low-cost system to classify squid in industrial production. This involved designing an easy-to-use handheld image-acquisition system combined with an automated, labor-saving, and efficient deep learning model (named “improved faster recurrent convolutional neural network”) to identify three squid species from the North Pacific Ocean. Three indicators, Accuracy, Intersection-over-Union, and Average Running Time, are used to evaluate the classification, and the average results for the test samples are 85.7%, 80.1%, and 0.144 s, respectively. The proposed network provides better squid classification compared with four other approaches. In addition, to ensure quality, the freshness of the selected squid is also evaluated using global threshold segmentation analysis. This proposed method is demonstrated to be a robust, noninvasive, and high-throughput system for squid classification and can also be expanded to other fine processing of aquatic products.