Multiple Color Feature and Contextual Attention Mechanism Based on YOLOX

Abstract

Object detection aims to find out and classify objects in which people are interested. YOLOX is the one-stage object detector representative, with being famous for its quick speed. Nevertheless, recent studies have illustrated that YOLOX suffers from small-scale accuracy. To deal with this issue, enlarging recent datasets and training models are adopted. It, however, results apply effectively on those big datasets, not on small or personal self-collected ones. It means a lack of datasets on special situations, like plastic runway surfaces. To enhance the performance on detecting small-scale objects, in this paper, we proposed Double-C YOLOX, an improved algorithm based on YOLOX. The model adds an HSV module and convolutional block attention to achieve more feature extraction. Due to the scarcity of plastic runway surface and the similarity of road damage, the published road dataset and selfie dataset are combined to train and test the performance of the proposed method. Experiments show that our model improves the mAP score by 2.82%. Double-C YOLOX is more suitable for detecting the small damages, such as hole and crack, than YOLOX.

This work is supported by National Natural Science Foundation of China Youth Fund (No. 61802247), Natural Science Foundation of Shanghai (No. 22ZR1425300) and Other projects of Shanghai Science and Technology) Commission (No. 21010501000). “Plastic Runway Surface Damage Detection System with YOLO” project for the Shanghai 2022 College Student Innovation Competition.