Study on the Plate Detection Method Based on the Data Generated by Significance Detection

EasyChair Preprint 12860

Abstract

In the smart restaurant environment, vision-based plate detection technology is the core of automated dish management. This study focuses on the identification of the existence of dishes on the plate, and constructs a dataset of dishes on the plate by collecting data from the actual restaurant scene. In order to complete the detection of dinner plates, a plate detection model based on Yolov5 was designed. Due to the limitation of the available restaurant dishes, the dataset has a single dish variety on the plate, which leads to the low performance of the trained plate detection model. In order to solve this problem, this paper proposes a plate data generation method based on saliency detection, which uses its category-independent characteristics to extract a variety of dish data from a variety of scenarios, and completes the annotation of the generated data by designing prompts and combining with Grounding DINO, which effectively solves the problem of small amount of plate data and single type. Experimental results show that the proposed plate detection model can effectively detect the presence of dishes on plates in a variety of restaurant environments, and the data generation method based on saliency detection significantly improves the quality of the dataset and the performance of the plate detection model. 

Keyphrases: Plate detection, Salient object detection, data generation, image editing

@booklet{EasyChair:12860,
  author    = {Linzhong Fang and Xuanlai Tang and Shuyon Gao and Yicheng Song and Shengfeng Li},
  title     = {Study on the Plate Detection Method Based on the Data Generated by Significance Detection},
  howpublished = {EasyChair Preprint 12860},
  year      = {EasyChair, 2024}}