Single-stage Multi-scale Receptive Field Improvement Lightweight Object Detection Network Based on MobileNetV3

Abstract

The computing power of edge devices is difficult to keep up with the development of modern computer technology, and the computing power is not improved enough. In practical application environments, so there is the birth of lightweight models. Lightweight models specifically refer to some models with simple model architecture and low computational load. Although the lightweight model is fast and the model is relatively simple, the detection effect is not very good. This paper proposes a parallel convolution module, performs feature fusion through parallel processing of multi-scale convolution kernels, and then integrates the spatial channel attention mechanism into the module to implement a multi-scale target detection module on a single feature layer. Model fusion proposes anchor boxes to generate heads and become a single-stage object detection model.