Study On Lightweight Convolutional Neural Networks Architecture And Its Application In Object Detection

High-performance convolutional neural networks often have many parameters,calculations,and complexity,and rely heavily on high-performance GPUs and other hardware,making it difficult to deploy convolutional neural network technology to embedded devices or mobile terminal equipments and other low-end devices with limited memory and computing capabilities.At the same time,the vigorous development of convolutional neural networks has accelerated the related research of object detection algorithms based on convolutional neural networks.As an important sub-topic in the field of object detection,pedestrian detection plays a role in intelligent monitoring,intelligent assisted driving,pedestrian analysis and other fields.important role.This paper studies the design of lightweight convolutional network architecture and the application of lightweight architecture in pedestrian detection tasks.The main work is summarized as follows:1?This paper studies the design method of lightweight convolutional network architecture and designs the lightweight convolutional neural network architecture C-Net.The method of cross-channel cross fusion is proposed.The two-stage dimensionality reduction method and the cross addition and fusion of the feature maps of adjacent groups are adopted to reduce the number of parameters and overcome the disadvantages of the lack of information exchange between different groups.The method of cross-module connection is proposed.Add shortcut connections between feature maps after dimensionality reduction of different basic building blocks within the same stage,which overcomes the disadvantage that the basic building blocks are independent of each other in the traditional lightweight architecture.Based on the two proposed methods,the lightweight convolutional neural network architecture C-Net is designed.The performance of the Food?101 and Caltech?256 datasets is compared with different network architectures.The experimental results verify the high performance of C-Net.performance.2?This paper studies the application of C-Net in object detection and designs a lightweight pedestrian detection model C-PDet(Pedestrian Detection Based on C-Net).The backbone network VGG-16 of the SSD(Single Shot Multibox Detector)detection algorithm was replaced with C-Net and fine-tuned.The channel attention mechanism was added to the position of the basic building block after cross-channel fusion dimensionality reduction,and the cross-module connection was further optimized into the cross-module attention connection.A sandwich module was designed to integrate the basic building blocks at the beginning and end of each stage to construct a more complex and comprehensive Feature map through the attention mechanism,and on this basis,a Compound Feature Pyramid Network(CFPN)was designed.A more refined setting of the priori box was made,and the result filtering policy was adjusted.Finally,the experimental results of C-PDet on the INRIA and Caltech pedestrian datasets show that C-PDet can better meet the requirements of real-time detection in terms of accuracy,speed and model volume.This paper has 44 figures,12 tables,and 102 references.