Improvement Based On Single Shot Multibox Detector Target Detection Model

With the presentation of Convolutional Neural Network(CNN),object detection based on convolution neural network has been applied in more and more fields.In the field of object detection,the theory of convolution neural network is maturing day by day.More and more excellent algorithms are improved on the basis of convolution neural network,such as Fastrcnn,Faster-rcnn,YOLO,SSD and so on.The SSD(Single Shot multibox Detector)model uses single depth neural network to detect objects in the image,and frames the detected objects and outputs the borders.These borders are generated on different structures in the SSD model,each with a different width to height ratio and size,so as to choose a box that is most suitable for the aspect ratio and size of the image of the current object.In the prediction stage,we should calculate the score of each object in the box.At the same time,the position of the border is fine-tuning to match the rectangle of the object.By combining different resolution feature maps,SSD can match the same object from multiple resolutions.The advantage of the SSD model is that the real time is good.However,the SSD model still has insufficient light and shadow processing,and can not accurately identify the problem of small and medium objects in the image.These problems restrict the accuracy of the SSD model to detect and identify objects.In this paper,an improved SSD model is proposed to improve the problems of SSD model.Improved SSD model calculation process: first,the image is preprocessed using HOG algorithm to reduce the impact of light on object recognition;then,on the basis of SSD model,the region recommendation algorithm and anti coiling layer are added,and the object and background can be quickly identified by using the regional recommendation algorithm in the object detection stage.In the phase of object recognition,the improved VGG algorithm can be used to solve the problem of the identification precision of the same object in different sizes,the size normalization of the object pixel and the low precision of the small object by adding the reverse coiling layer.Finally,the 5 tasks of changing into the SSD model training in the model training stage are described and used.Natural random gradient algorithm reduces the error in training.This experiment includes the effect of the object detection algorithm and the size of the training set on the detection precision of the improved SSD model,and the corresponding experimental data are tested on different data sets.The YOLO and SSD object detection algorithms are compared,and the data are obtained through experiments,and the experimental results are analyzed.Finally,the software of object detection based on improved SSD model is designed,and the software is tested and displayed.To sum up,the improved SSD model reduces the impact of light on the picture,adding anti coiling layer on the original SSD model infrastructure to improve the detection precision of small objects,and reduce the probability of detection of the same object by multiple detection frames,increase the operation efficiency,and be more in line with the requirements in the actual application.