Real-time Object Detection Based On Multi-scale Neural Network And Self-attention Mechanism

With the rapid development of science and technology and the rapid growth of data and information,the demand for intelligent image processing technology in various fields is increasing rapidly.Among them,the most important branch is object recognition and detection based on deep learning.Because of its fast processing speed,wide detection range,and the ability to process large amounts of data at the same time,it is widely used in intelligent video surveillance,pedestrian re-identification,enemy ship detection,marine rescue monitoring and other civilian and military fields.At present,the object detection algorithm based on deep learning has made breakthrough progress,and the detection performance has also been greatly improved.However,in many application scenarios that require higher real-time performance and accuracy,the existing real-time object detection system still has the disadvantages of bulky and complicated network,low detection rate,and low accuracy of object detection at different scales.In this paper,deep convolutional fusion network,multi-scale feature fusion module based on pyramid,and embedded module based on self-attention mechanism are applied to real-time object detection tasks,so that the network can detect objects with different shapes and sizes in both accuracy and speed.There has been a clear improvement.The main research content of this paper includes the following aspects:1.A object detection method based on deep convolution fusion network is proposed.Aiming at the disadvantage of the unbalanced semantic features of different positions in the deep convolutional network and the large difference in detection accuracy of objects of different scales,a multi-scale feature fusion module based on a pyramid structure is added to the basic network.Aiming at the disadvantages of the unbalanced positive and negative sample frames in the deep convolutional network and the inability to accurately locate the object and the background,a multi-scale feature fusion module based on a two-stage structure is added to the basic network.Finally,on the VOC data set,the deep convolutional network and the multi-scale feature fusion network are tested and compared with the results.It can be found that the detection accuracy of the deep convolutional fusion network has been greatly improved compared with the basic convolutional network.2.A object detection method based on pyramid multi-scale neural network is proposed.Aiming at the single feature extraction method of the SSD network,which leads to the disadvantage of being unable to accurately locate small objects,dense objects,and occlusion objects,a pyramid-based feature extraction method is added to the SSD network.This enables the network to simultaneously extract and merge the position offset and global contour features of the object.Finally,the pyramid multi-scale object detection network is experimentally tested on the VOC data set and the COCO data set,and compared with the detection accuracy of the excellent network in recent years,which fully shows the superiority of this network.3.An embedded object detection method based on self-attention mechanism is proposed.Aiming at the disadvantage of the feature extraction method of the SSD network that is too blind and the feature information is too redundant,the recognition accuracy of objects of different scales is low,a prediction module based on the self-attention mechanism is added.Aiming at the problem of deep convolutional network operation speed is too slow and poor portability,the multi-scale neural network based on self-attention mechanism is loaded on FPGA,using the low power consumption of FPGA and fast operation speed to fully realize the network model acceleration optimization.Finally,on the VOC data set,the detection accuracy and detection rate of the self-attention embedded object detection system were experimentally tested and compared with the results,which fully showed the superiority of this system.In summary,this paper studies the real-time object detection method based on multi-scale neural network and self-attention mechanism.The experimental results prove the feasibility and effectiveness of the proposed method,and show that the research results have certain practical significance.