Deep Learning Based Object Detection Technology Research Under Android Mobile Platform

Object detection is the core and basic research issue in the field of computer vision,which has been widely used in video surveillance,biomedicine,and automatic driving.Compared with the traditional object detection algorithm,the object detection algorithm SSD(Single Shot Multi Box Detector)based on deep learning has greatly improved the detection accuracy.The object detection algorithm SSD use the VGG-16 convolutional neural network to extract features.The VGG-16 has many weight parameters and a large amount of calculation,which limits the application of SSD object detection algorithm on mobile phones,embedded devices and other limited-performance devices.The purpose of this paper is to reduce the computational complexity of the SSD object detection algorithm so that it can have a wider range of applications.(1)Mobiel Net is a lightweight convolutional network for mobile and embedded devices.The compression method is to calculate the L1 norm of the Mobile Net pointwise convolutional convolutional filter,and remove several convolution filters with the smallest L1 norm.The compress effect is: The size of Mobel Net model is reduced by 18.6% compared to Mobile Net prototype parameters.And The compression accuracy of Mobile Net Top-1 and Top-5 is 65.2% and 85.8%,respectively.(2)The compressed Mobile Net replaces the VGG-16 as the basic network architecture of the SSD object detection algorithm.And eight convolutional layers are added to the tail of the basic Mobile Net.Six different scales of feature maps are used for convolution operation to detect targets.The PASCAL VOC2007 data set is selected to train and test the improved SSD detection model.The experiment hardware environment is CPU i7 GPU GTX850 M,and the software environment is Tensor Flow deep learning platform.The mean average precision m Ap of the improved SSD detection algorithm is 67.4%,and the average target detection time of the improved SSD object detection algorithm is 19.42 ms.In the same hardware test environment,the average detection time of the original SSD prototype is 118.96 ms.Compared with the SSD prototype,the average detection speed of the improved SSD is increased by nearly 6 times.(3)The trained object detection model is transplanted to the Android mobile operating system and the Android application program of the target detection is written.The application of Android object detection is used to test the performance of the improved SSD object detection model on mobile devices,which the hardware configuration of the mobile device is the CPU Hisilicon Kirin 950,4GB of running memory.The test results are as follows: The Android application of object detection occupies the memory of mobile phone between 110 M and 130 M,and The CPU usage is 25% and 45%.Moreover,the target detection time is between 1s and 3s.Compared with the original SSD target detection algorithm,the improved target detection algorithm SSD in this paper has less computational volume and can run smoothly on mobile devices.Moreover,the improved SSD algorithm has better object detection effect and cost less detection time,which shows it has practicality.