Light-weight Object Detection Algorithm Based On Multi-scale Structure

In recent years,domestic and foreign scientific research institutions,universities and enterprises have gradually paid attention to and developed the potential value that artificial intelligence can bring,and computer vision is one of the core scenarios of artificial intelligence applications.Object detection is an important branch of computer vision.Its task is defined to determine the type of objects contained in an image and their position in the image.The research work of object detection can serve many applications in real life,such as filtering out illegal and bad network information through computer vision technology such as object detection,thereby improving the efficiency of cyberspace governance.In addition,it also has application value in scenarios such as autonomous driving and assisted medical care.This research is divided into two parts,exploring the lightweight object detection algorithm and feature fusion module respectively,and proposes a lightweight object detection algorithm based on multiple receptive fields and a object detection algorithm based on multiple rounds of feature fusion.The main tasks include:First,analyze the problems existing in the construction of lightweight neural networks in the past and propose corresponding optimization directions.A set of depthwise separable convolutions with different convolution kernel sizes and conventional convolutions with a convolution kernel size of 1 are used for feature extraction,combined with a simplified feature fusion module to alleviate the decrease in accuracy due to the smaller neural network scale.Based on the above ideas,a lightweight object detection network LME SSD is constructed.The experimental results show that the parameter amount of LME SSD is 5.85M,the computational complexity is 2.20B,and the mAP value on the data set Pascal VOC2007 is 72.8%.In terms of accuracy,LME SSD surpasses lightweight networks such as MobileNet.In addition,its parameter amount and computational complexity are far lower than non-lightweight object detection networks.Secondly,in the past,the construction of feature fusion modules in the object detection network has the problems of excessively high computing resources and insufficient feature fusion.Follow-up research work is carried out to solve the above problems.The depthwise separable convolution with different convolution kernel sizes is combined with bilinear interpolation to complete the sampling operation,thereby reducing the computational cost in the module construction,and ensuring the adequacy of feature fusion by performing multiple rounds of information fusion on the feature map.Finally,the designed feature fusion module is grafted on SSD(HLF SSD)and FSSD(HLF FSSD)to run experiments.Experiments show that the mAP values of HLF SSD and HLF FSSD on the Pascal VOC2007 data set are 79.2%and 80.0%,which exceed the accuracy of SSD and FSSD,and the loss of detection speed is small.