Object Detection Based On SSD Mobilenet In Deep Learning

Target detection is an important research field in computer vision and image processing.With the advent of the era of big data,the entire industry and academia are using deep learning for target detection.Although the target detection accuracy based on deep learning is far more than the traditional target detection method,it still falls far short of the requirements of human beings.In particular,some algorithms are even lower than the traditional target detection algorithm in the detection speed,so it is difficult to achieve real-time detection in many target detection in deep learning.In order to speed up the detection speed of the model,this paper mainly explored the SSD_Mobilenet model through the combination of SSD(Single Shot MultiBox Detector)and lightweight network Mobilenet.Compared with SSD,this model greatly improves the detection speed,but the detection accuracy still needs to be improved.In order to improve the detection accuracy of SSD_Mobilenet,the following is our main contributions:Aiming at the problem of insufficient positioning information in SSD_Mobilenet,a large number of false detection frames and low detection accuracy of occluded objects,a multi-scale semantic information fusion model is designed,which can effectively reduce the loss of information after convolution and improve the accuracy of targetdetection.In the open benchmark VOC 2007,SSD_Mobilenet was 1.4%higher than the the improved SSD_Mobilenet mAP.In order to improve the detection accuracy of single object,this paper presents to analyze the data in the training data set,the shape and size of the target real frame and its proportion in the whole image.According to the data analysis,a set of reasonable aspect ratio of anchor points and the number of anchor points frames of each anchor point are designed.Experimental results show that our method improves the detection accuracy of single target object and can locate the target more accurately.