FD_YOLOX: An improved YOLOX object detection algorithm based on dilated convolution

Abstract

In autonomous driving perception, vehicle object detection based on deep learning has been a major research topic. However, detecting vehicles has been a complex issue in computer vision. To address this challenge, we present the FD_YOLOX algorithm with three effective modules. The Fusion Dilated Convolution Module (FDCM) is combined dilated convolution feature fusion with an attention mechanism, which is allowed the network to obtain richer semantic information while adaptively detecting targets. The Dilated Channel-Adjusted Convolution (DCAC) is also proposed to address the small receptive field of high feature layers and adjust the number of channels in the input feature layer. Finally, the Dilated Spatial Pyramid Pooling (DSSP) is built by introducing dilated convolution in SPP, enhancing the receptive field of the network and preserving more information about small targets and their locations. Through experiments on the SODA10M data set, the FD_YOLOX_s model showed an increased mean average precision of 2.05%(640× 640) and 1.15%(1280 × 1280) compared to YOLOX_s. Moreover, the FD_YOLOX model achieved competitive performance against other advanced object detection algorithms in the detection of small target vehicles.