Object Detection Algorithm Based On Multi-layer Convolution Features

Object detection is widely used in industry,agriculture and real life.There are many factors that affect the detection of algorithms in real production and life scenes,such as occlusion and background interference of the target to be detected and illumination variations.If the accuracy of the detection algorithms is not high or cannot meet the real-time requirements,the algorithms are difficult to put into practice.Traditional object detection algorithms design features manually according to the characteristics of the target to be detected,and then train appropriate classifiers to classify these features.For object detection,traditional detection algorithms are difficult to design features that are suitable for all targets,so the algorithms are often get poor accuracy and the detection speed is low.Aiming at the above problems,object detection algorithms based on deep learning learn the features of all kinds of targets to be detected by deep convolutional neural networks,and then use the learned features to detect target of various categories and the detection speed is comparatively faster.This paper analyzes the popular object detection algorithms at this stage,thoroughly studies the correlative theories of deep convolutional neural networks,improves the existed network models,and designs a new network model which is suitable for object detection.The main works of this paper are as follows:Firstly,aiming at the problem that the accuracy of SSD convolutional neural network model is not high for small object detection,an improved SSD model based on feature pyramid network is proposed.The feature pyramid network can fuse the deeper convolutional feature maps which have more abstract and richer semantic information and the shallower convolutional feature maps with higher resolution and more detailed information.This is very important to the detection of small targets.The detection process is that multi-layer feature maps obtained from the original SSD network are processed to obtain feature maps of fusion layer by the lateral connection layer,upsampling layer,fusion layer,and prediction layer all of which are designed by the paper,the feature maps of fusion layer are used to perform object detection after processing by the prediction layer,and then the final detection result is obtained by non-maximum suppression.Secondly,taking the classical VGG model as the basic network framework,adding batch normalization,Dropout mechanism,dense connection and other skills to improve the network model,and then combine the upsampling and channel splicing techniques to realize the object detection based on multi-layer convolutional feature.Among them,batch normalization can keep the same probability distribution for each round of data input of convolution kernel;Dropout can make the network connection of convolution kernel more diverse during training,and dense connection can increase the reuse of shallow high-resolution feature maps.Experiment results in the PASCAL VOC show that the improved network has a significant improvement in accuracy through the multi-layer convolutional features prediction mechanism.Finally,the training skills used in the experiments and encountered problems in pre-training model are analyzed and summarized.Combining the classic object detection models SSD and VGG,we propose an end-to-end detection algorithm based on deep learning,which improves the accuracy of multi-class object detection algorithm to a certain extent.