SR-YOLO: Small Objects Detection Based on Super Resolution - IFIP Open Digital Library
Conference Papers Year : 2022

SR-YOLO: Small Objects Detection Based on Super Resolution

Abstract

Since the introduction of convolutional neural networks, object detection based on deep learning has made great progress and has been widely used in the industry. However, because the weak and small object contains too little information, the samples are rich in diversity, and there are different degrees of occlusion, the detection difficulty is too great, and the object detection has entered a bottleneck period of development. We firstly introduce a super-resolution network to solve the problem of small object pixel area being too small, and fuse the super-resolution generator with the object detection baseline model for collaborative training. In addition, in order to reinforce the weak feature of small objects, we design a convolution block based on the edge detection operator Sobel. Experiments show that proposed method achieves mAP50 improvement of 2.4% for all classes and 4.4% for the relatively weak pedestrian class on our dataset relative to the Yolov5s baseline model.
Fichier principal
Vignette du fichier
537972_1_En_38_Chapter.pdf (8.32 Mo) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-04666410 , version 1 (01-08-2024)

Licence

Identifiers

Cite

Biao Hou, Xiaoyu Chen, Shenxuan Zhou, Heng Jiang, Hao Wang. SR-YOLO: Small Objects Detection Based on Super Resolution. 5th International Conference on Intelligence Science (ICIS), Oct 2022, Xi'an, China. pp.352-362, ⟨10.1007/978-3-031-14903-0_38⟩. ⟨hal-04666410⟩
22 View
2 Download

Altmetric

Share

More