Research on lightweight real-time object detection based on attention mechanism

Image object detection is a fundamental and challenging task in the field of computer vision. Its goal is to detect the smallest border covering the region of interest and assign the corresponding semantic label to the border. In recent years, with the development of deep learning, the performance of target detectors has been greatly improved. Although these detectors have made great strides, most of them contain hundreds or even thousands of convolutional layers and feature channels whose model size and execution efficiency are unacceptable for real-world applications that require real-time inference, such as autonomous driving and machine vision. In order to adapt to real-world applications, a number of lightweight real-time target detectors have been proposed, which typically employ a single-branch backbone structure, contain a small number of convolutional layers and continuous downsampling operations, which makes their high-resolution positioning details rough and inaccurate, which is unfavorable for target positioning. On the other hand, due to the limited model capacity of lightweight networks, they are not good at modeling global relationships. In addition, in order to obtain faster inference speed, existing lightweight models tend to use a simple architecture in the neck and detection head of the detector, but this practice ignores the relationship between different features. In order to deal with the above problems, this paper conducted the following research: To solve the problem of insufficient high-resolution detail extraction due to the fast downsampling strategy of lightweight detectors in shallow layers of the network, a dual-path network (DPNet) with lightweight self-attention is proposed for efficient object detection. DPNe uses a dual-path backbone architecture, which has two parallel dual-resolution paths, in which the low-resolution path extracts semantic information and the high-resolution path retains positioning details, both of which are very important for object detection. In addition, to improve the capacity and feature representation of the model, a lightweight Self-attention module (LSAM) with single input and single output is designed and embedded in the backbone network. Experimental results in MS COCO 2017 show that DPNet achieves a satisfactory balance between detection accuracy and execution efficiency.