Research On Single-Stage Object Detection Algorithm Based On Deep Convolutional Neural Network

With the rapid development of big data technology and the country's strong support for the field of artificial intelligence,object detection technology based on deep learning is widely used in pedestrian detection,face detection,driverless cars,smart cities and express logistics.At the same time,the application of the depth convolution neural network in object detection technology can automatically extract the rich characteristic information of the image,which well solves the poor generalization ability of the network in the traditional target detection algorithm,the manual design features are difficult to resolve such problems as the target diversity,and the detection speed and accuracy have great improvement.However,the object detection technology based on deep learning cannot guarantee the detection accuracy while have a fast detection speed.Therefore,we need to better balance the relationship between detection accuracy and detection speed in the target detection algorithm.As a representative single-stage target detection model,YOLOv3 perfectly balances the relationship between detection accuracy and detection speed,which has high reference value in practical engineering applications.In the paper,based on the depth of the convolution of single stage object detection algorithm of neural network research direction,combined with the existing YOLOv3 single stage detection algorithm,we design a lightweight backbone network with Attention mechanism,the design of the bounding box of the metrics,converges more rapidly accurate loss function design,and in the prediction stage to a non-maximum suppression algorithm improvement measures such as put forward a new kind of single stage object detection algorithm,concrete measures are as follows:(1)In order to further improve the detection speed of YOLOv3,in the process of network model design,we use depth separable convolution to replace the traditional convolution operation,so that the network can effectively extract image features while greatly reducing network parameters.At the same time,the algorithm introduces attention mechanism in space and channel in each convolution module of the network,which makes the network pay more Attention to the target information in the image and extract complex image feature information more effectively.(2)In the process of model training,we redesigned the measurement standard between theprediction boundary boxes,used GIo U algorithm to more accurately represent the overlap degree of the two boundary boxes,and redesigned the original loss function in combination with GIo U.Concrete,in order to speed up the convergence of loss function,this article uses the degree of overlap between the bounding box and the center of the distance between two bounding box as the loss function instead of the original computing center of loss function and boundary width and height loss function,and use YOLOv3 loss function,the original object categories and confidence of the bounding box in the training process has a faster and more accurate regression,also makes the detection algorithm for small object detection in image is more friendly.(3)In the process of model prediction,the original non-maximum suppression algorithm was improved by combining with GIo U algorithm,gaussian model was used to suppress the surrounding boundary boxes instead of deleting them when screening candidate boxes,and the value of the distance between GIo U and the boundary box center point was calculated as the threshold for screening candidate boxes.This enables the model to avoid deleting occlusion targets in dense images to a certain extent,and to screen out more accurate candidate boxes.Finally,under Pytorch framework,MSCOCO2017 dataset is used to train and test the improved object detection model.Experimental results show that,compared with the original YOLOv3 detection model,the single-stage object detection model designed in this paper performs slightly worse in the overall detection accuracy and the detection accuracy of larger objects or single objects in the image,but performs better in the detection of small targets and the occlusion of targets.At the same time,compared with other mainstream target detection algorithms,the single-stage detection model designed in this paper has a significant improvement in detection speed.In addition,in order to further improve the universality of the single-stage detection algorithm designed in this paper,the paper uses KAIST multi-spectral pedestrian data set to train and test the designed model and detect pedestrians in infrared images.The test results show that the object detection algorithm designed in this paper also has good detection results in infrared image detection.