ST-CenterNet: Small Target Detection Algorithm with Adaptive Data Enhancement

Abstract

General target detection with deep learning has made tremendous strides in the past few years. However, small target detection sometimes is associated with insufficient sample size and difficulty in extracting complete feature information. For safety during autonomous driving, remote signs and pedestrians need to be detected from driving scenes photographed by car cameras. In the early period of a medical lesion, because of the small area of the lesion, target detection is of great significance to detect masses and tumors for accurate diagnosis and treatment. To deal with these problems, we propose a novel deep learning model, named CenterNet for small targets (ST-CenterNet). First of all, due to the lack of visual information on small targets in the dataset, we extracted less discriminative features. To overcome this shortcoming, the proposed selective small target replication algorithm (SSTRA) was used to realize increasing numbers of small targets by selectively oversampling them. In addition, the difficulty of extracting shallow semantic information for small targets results in incomplete target feature information. Consequently, we developed a target adaptation feature extraction module (TAFEM), which was used to conduct bottom-up and top-down bidirectional feature extraction by combining ResNet with the adaptive feature pyramid network (AFPN). The improved new network model, AFPN, was added to solve the problem of the original feature extraction module, which can only extract the last layer of the feature information. The experimental results demonstrate that the proposed method can accurately detect the small-scale image of distributed targets and simultaneously, at the pixel level, classify whether a subject is wearing a safety helmet. Compared with the detection effect of the original algorithm on the safety helmet wearing dataset (SHWD), we achieved mean average precision (mAP) of 89.06% and frames per second (FPS) of 28.96, an improvement of 18.08% mAP over the previous method.