Research On Light-weighted And Anchor-free Object Detection Methods

As a basic task in the field of computer vision,object detection aims to locate and classify objects of interest in an image,which has a wide range of application scenarios across many fields,such as automatic driving,image segmentation,and text detection and recognition.In scenarios such as autonomous driving and the Internet of Things,where there are strict requirements for real-time algorithm performance and low power consumption of computing devices,it is necessary to limit the number of parameters and computational complexity of the detection model.Hence,research on lightweight object detection models is crucial for efficient algorithm implementation.This thesis proposes a lightweight object detection model that focuses on the network structure and pruning methods to achieve a more efficient solution.The main research contents of this thesis are as follows:1.A lightweight method of detection head suitable for anchor-free detection model is proposed in this thesis.This thesis analyzes the computational cost of a typical anchorfree detection model.For the detection head structure with the most computational cost,a method is proposed to reduce the computational cost of the stacked convolution component by reducing the spatial resolution of the input feature map.An information loss-less up-sampling layer is added before the predictor structure to restore the spatial resolution of the feature map,so as to ensure that the small objects’ detection ability of the model is not reduced to the greatest extent.This lightweight detection head method can reduce the FLOPs of the original detection head to about half,while keeping the detection performance almost unchanged.2.A light-weighted multi-scale feature fusion module with visual attention applied to feature pyramid network is proposed in this thesis.This thesis designed a lightweight visual attention module for multi-scale feature fusion by combining the existing convolutionbased channel domain and spatial domain attention modules and self-attention mechanism,and integrating the computational saving methods used in them.In addition,this thesis combines the concatenation-based feature fusion mechanism and the attention-based mechanism to form a lightweight multi-scale feature fusion module.This module can reduce the FLOPs and parameters of the detection model under the premise of maintaining the main performance metrics of object detection.3.This thesis focuses on the processing of pruning importance scores based on moving average methods.Aiming at the contradiction between the total time cost of pruning and the pruning effect in the existing pruning methods based on gradient information,this thesis proposed to use the exponential moving average method to smooth the importance score calculated by multiple iterations.On the one hand,the moving average method can increase the number of samples that can be considered in a single pruning operation while maintaining the overall pruning interval.On the other hand,the weighting coefficient of exponential moving average can simultaneously consider the difference of model structure before and after pruning operation,and the influence of recovery training on the accuracy of importance score calculation.This importance score processing method can improve the final pruning effect at a lower time cost.