Research On Improved SSD Target Detection Algorithm Based On Transposed Convolution Operation

Object Detection refers to the technology that automatically determine the location and the type of objects in images using computer,which is also one of the key technologies in the fields of security,auto-driving and astronomy.However,since the objects in images vary in quantity,position and size,object detection algorithms are difficult to implement.There is no recognized optimal method in the field,so it is meaningful to further to improve accuracy and performance of current object detection algorithms.With dramatic improvement of computer performance and massive data generated by the Internet,deep learning has also ushered in a stage of development.Currently,all of the fastest and most accurate algorithms are obtained by using deep learning method.In order to further improve performance of existing object detection algorithms,the thesis has carried out following studies:(1)Exploring the impact of various factors on performance of object detection algorithms.All modern object detection algorithms use convolutional neural network as fundamental component.However,it can be difficult for practitioners to decide which is best suited to their specific application.Object detection algorithms are divided into two categories: single-stage and two-stage methods.Single-stage methods have high accuracy,but are computationally expensive;Two-stage methods satisfy the real time requirement,but are not as accurate as single-stage methods.When put algorithms into application,running time and memory usage are also critical.So,in this thesis,the effects of feature extraction networks,size of input images and other factors on the detection accuracy,memory footprint,FLOPs are analyzed experimentally.Key characteristics of typical object detection algorithms are summarized.(2)Based on the analysis and comparison results of typical object detection algorithms,a recurrent feature aggregation model was proposed based on SSD and applied to vehicle detection scenarios.Since the mean Average Precision(m AP)of SSD drops significantly when evaluating with higher Intersection over Union(Io U),a feature aggregation method using transposed convolution as main component was proposed.On the basis of SSD,a deep Residual convolutional Network(Res Net)with 101 layers was used to extract features.Firstly,abstraction of semantics and context information was generated using transposed convolutional layers which double the scales of deeper feature maps.Secondly,fully connected convolutional layers were applied to shallow layers to prevent unexpected bias.Finally,the shallow and deep feature maps were concatenated together,and convolutional layers with kernel size 1 were used to reduce the channel sizes.The feature aggregation can repeat multiple times.The experiments were conducted on KITTI dataset and took 0.7 as Io U threshold.Experimental results show that the m AP was improved by about 5.1 percent points compared to the original SSD model,about 2 percent points compared to the state-ofthe-art Faster R-CNN model,and maintained the original level of running speed of SSD.The feature aggregation model can effectively improve the m AP and generate high quality bounding boxes in object detection tasks.