Research On Neural-Network-Based Lightweight Algorithms For Objection Detection

Object detection is an important and fundamental task in the field of computer vision,which typically serves as the basis for downstream visual tasks.Over the past few years,as deep learning has continued to advance,many object detection algorithms have been able to achieve excellent accuracy.However,these large and complex network models require significant computing resources.Therefore,the focus of this study is to improve existing network frameworks in the field of object detection,with the aim of maintaining model detection accuracy while lightening the neural network.This dissertation aims to accurately and efficiently detect objects,and proposes two lightweight object detection algorithms based on depth-wise separable convolution and attention enhancement,as well as based on structural re-parameterization.(1)A lightweight object detection algorithm based on depth-wise separable convolution and attention enhancement is proposed:This dissertation optimizes the network structure of the anchor-based one-stage YOLOv3 to achieve a lightweight object detection algorithm based on depth-wise separable convolution and attention enhancement.The optimization is mainly carried out from the following two aspects: In terms of lightweight,this dissertation reconstructs the network backbone based on MobileNetv2,replacing each 3x3 convolution in the neck and head of the object detection network with a fusion structure of 3x3 depth-wise separable convolution and compressed excitation module;In terms of attention enhancement,the high-frequency wavelet of the original image is used as an additional channel input,and a simplified non-local attention module is added to the appropriate position of the simplified backbone.The convolutional block attention modules(CBAM)are then introduced to the simplified neck of the object detection network.In addition,the local 3x3 convolution branches across multiple module sequences are added to the simplified backbone network to enhance the learning ability of the network.The experimental results show that this method outperforms existing lightweight object detection algorithms in one or more evaluation aspects,especially in terms of the average precision(mAPs)of detecting small objects(area < 32~2)(area is measured as the number of pixels in the segmentation mask).(2)A lightweight object detection algorithm based on the idea of structural re-parameterization is proposed:This dissertation provides a reasonable analysis and light-weighting strategy of the network structure of the anchor-free one-stage NanoDet-Plus.For the ShuffleNetv2 backbone of the NanoDet-Plus object detection network,in order to make it more suitable for object detection tasks.This dissertation proposes a novel lightweight object detection algorithm based on the idea of structure re-parameterization,which main contributions can be divided into three parts: Firstly,a novel lightweight channel attention module was proposed to enhance the model accuracy.Meanwhile,convolutional kernel dilation was applied at appropriate locations in the network.Secondly,in order to reduce the network complexity,the compact Ghost module was utilized to replace the 3x3 standard convolution.Lastly,to reduce the inference time consumption of the model,the training and inference stages of the network were decoupled and the structural re-parameterization idea was incorporated into the entire network.The experimental results show that this method performs better than the baseline method in multiple average precisions while reducing both the parameter and computational complexity.Additionally,in one or more evaluation metric of total parameter,computational complexity,and average precision,it outperforms all compared advanced benchmark algorithms.