Single Shot Object Detection Algorithm Based On Feature Fusion And Enhancement

Object detection is a main research domain in computer vision,and it has many practical applications,such as unmanned driving,intelligent medicine and so on.Traditional object detection algorithms are mainly based on the features extracted manually and the shallow structure that can be trained.In recent years,with the rapid development of deep learning technology,more and more researchers attempted to solve computer vision tasks with deep learning methods.Because deep learning can extract deeper,higher semantic levels and more robust features,object detection algorithms based on deep learning have made breakthroughs in this domain.However,many complex application scenarios not only require high accuracy of object detection,but also require real-time detection speed.Therefore,how to improve the accuracy and speed of the object detection algorithm further is still the focus of researchers.Aiming at the difference between object detection and image classification,and the problem that most object detectors depend too much on classification network,this thesis proposes a backbone network for object detection,called detection network(DNet).Image detail features are extracted using a small size convolution kernel in the initial portion of the network in the initial part of the network.In order to enhance the ability of the neural network to detect different scale objects,a feature fusion module is proposed to fuse the features from different convolution layers.In order to balance the ability of the neural network to locate and classify objects,a mix down-sampling module is introduced to make the features more robust after down-sampling.Experimental results on PASCAL VOC dataset show that DNet achieves a good balance between detection accuracy and efficiency compared with other backbone networks.In order to solve the problem that the object detector SSD has poor detection accuracy for small objects,the Pixel Shuffle up-sampling method is used to up-sample the deep feature map and combine with the shallow features to enhance the shallow features while hardly increasing the parameters of the network.In training process,labels are used to introduce the Assisted Excitation module.Because this module only exists in training,this method can improve the accuracy without affecting the test speed.Experiments on PASCAL VOC dataset show that the proposed method can improve the detection accuracy without reducing the detection speed.