Similarity Retrieval Based on Image Background Analysis

Abstract

Aiming at the problem of traditional portrait background similarity retrieval methods being low accuracy and time-consuming, a similarity retrieval method based on image background analysis is presented. The proposed method uses a combination of portrait segmentation and retrieval models. Firstly, the portrait segmentation model is used to remove the portraits in the images to eliminate the interference of portraits on background features; secondly, the image retrieval model is used to retrieve images with similar background features; LSH is added to improve the retrieval efficiency; finally, the retrieval results are used to further determine whether the background is similar. The experiment is implemented based on real data from a company. The results showed that the average precision, average map, and recall of this method reached 85%, 90%, and 50%, respectively. The average accuracy and recall are 10% better than the overall image retrieval model.

Unmanned Aerial Vehicle Fire Detection Platform Based on Semantic Yolov5 and Autonomous Recognition

“…We can solve this problem by adding images of unmarked flame objects as negative samples and further improve the model's performance. The author who created YOLOv5 suggests using a background image of approximately 0-10% to help reduce false positives (Zhu et al, 2022). The results of fire detection are shown in Figure 15.…”