Research On Target Detection Model Based On Mlti-level And Multi-scale Feature Fusion

Object detection is an important part in the field of image processing and has a wide application prospect in production and life.The core task of object detection is to identify the category and position of the target in the image.However,in daily images,there are huge scale differences between different targets,how to accurately identify all large-scale targets and small-scale targets is a challenging problem for object detection.Based on this,the multi-scale problem in object detection is studied in this paper.The main work of this paper is as follows:(1)To solve the problem of insufficient multi-scale feature information in object detection,the object detection model MLPNet based on convolutional neural network is proposed by using multi-scale and multi-level features containing multi-level information.The model uses multi-scale information fusion module to fuse deep and shallow features,uses feature reuse module to enhance the reuse and transmission of features between different levels,and the attention module is used for global context modeling and feature channel correlation,multi-scale features containing multi-level information are generated,the multi-scale feature fusion between different levels is strengthened,and the detection accuracy of large target and small target is effectively improved.The experimental results on COCO data set show that the AP of MLPNet reaches 37.9.Compare with M2 Det,the increase of the experimental results using the MLPNet model in the detection of small targets is obvious,reaching 4.8%,and the increase of the detection of large targets is 4.5%.(2)To solve the problem that features lose part of valid information after multiple convolution,multiple dilation convolution branches are used to form a dilation convolution module,and features are extended through the dilation convolution module for fusion.This module uses a parallel approach to extend features through dilation convolution with multiple different sampling rates and global average pooling layers,and the original feature information is transmitted through the residual connection branch,which can effectively solve the problem of information loss,and enhance the perception ability of the model to targets of different scales while retaining the original feature information.(3)To improve the accuracy of the object detection algorithm and the ability of multi-scale object detection,the multi-branch parallel dilation convolution module is used,and the D-MLPNet model is proposed based on the MLPNet model.In this model,the dilation convolution module is added to the multi-scale information fusion module to form the multi-scale information fusion dilation module.While retaining feature information,the multi-scale feature fusion among different levels is strengthened,which effectively improves the detection ability of the model for multi-scale targets.The experimental results show that the D-MLPNet model is better than the MLPNet model,and the detection accuracy for small targets is improved by 1.55% and that for large targets by 0.75%.