Improvement of SSD to Discriminate Small Foreign Materials in Fishes Product Line and Evaluation of Practicality

Abstract

The purpose of this research is to detect the location of foreign material in foods on factory line by the technique of object detection using deep learning. In a real factory line, both detection speed and detection performance are important. The proposed system is based on SSD (Single Shot Multibox Detector) which is known to be a relatively high-speed and high-performance. The proposed model improves the location detection performance for small foreign materials by increasing the size of the input image and modifying the SSD network structure. In order to evaluate the proposed model, the authors compared it with existing models. The experimental results show that the proposed model achieved 0.972 with F-value and its processing time was less than 0.35 seconds.