YOLO V4 with hybrid dilated convolution attention module for object detection in the aerial dataset

Abstract

ABSTRACT Object detection is an important part of computer vision. Besides, small object detection is a challenging task in object detection. Most existing methods have difficulty locating small objects and classification. In this paper, we propose a new method to solve the problem. On the one hand, we improve the YOLO V4 network with ASPP. On the other hand, we propose a Hybrid Dilated Convolution Attention (HDCA) module to focus on the important position in images. The Hybrid Dilated Convolution (HDC) module is redesigned for parameter-efficient in the HDCA module. We also design a Translational Dilated Convolution (TDC) to solve the ‘gridding issueʻ of the HDC and enlarge the receptive field at the same time. The experiments are based on the DOTA dataset, and our method achieves 2.31% mAP improvement compared with the original YOLO V4. Besides, our method achieves the best improvement in the class of basketball court, which reaches 81.03% of AP. Compared with the state-of-the-art method, our method achieves a 3.99% improvement on the mAP criterion. We put other attention modules in the YOLO V4 architecture at the same place as our method. And our method achieves 0.78% mAP improvement compared with the BAM module, which is the second place in the competition.