Abstract

At present, object detectors based on convolution neural networks generally rely on the last layer of features extracted by the feature extraction network. In the process of continuous convolution and pooling of deep features, the position information cannot be completely transferred backward. This paper proposes a multiscale feature reuse detection model, which includes the basic feature extraction network DenseNet, feature fusion network, multiscale anchor region proposal network, and classification and regression network. The fusion of high-dimensional features and low-dimensional features not only strengthens the model's sensitivity to objects of different sizes but also strengthens the transmission of information, so that the feature map has rich deep semantic information and shallow location information at the same time, which significantly improves the robustness and detection accuracy of the model. The algorithm is trained and tested in Pascal VOC2007 dataset. The experimental results show that the mean average precision of the objects in the dataset is 73.87%. At the same time, compared with the mainstream faster RCNN and SSD detection models, the mean average precision of object detection algorithm based on DenseNet is improved by 5.63% and 3.86%, respectively.

Highlights

  • Traditional object detection algorithms mainly include object feature extraction, object recognition, and object positioning

  • Histogram of Oriented Gradient (HOG) features are widely used in image recognition, especially for pedestrian detection, and a large number of optimization and improvement algorithms have been proposed successively. e detection algorithm based on the Deformable Part Model (DPM), proposed by Felzenszwalb in

  • In order to ensure the effectiveness of comparison, all algorithms are trained under the same hardware platform and software framework. e datasets used are all VOC2007 datasets, and the initialization and training mechanism used in Section 3.2.1 are used for the training of the four algorithms. e experimental results of the four algorithms are shown in Table 4. e classical target detection algorithms fast RCNN, faster RCNN, and SSD (300) achieved 68.24%, 70.01%, and 71.19% MAP. ey are 5.63%, 3.86%, and 2.8% lower than the algorithm proposed in this paper

Read more

Summary

Introduction

Traditional object detection algorithms mainly include object feature extraction, object recognition, and object positioning. Its process can be expressed as the extraction of image deep level features and the object classification and regression based on convolution neural network [4]. With the continuous subsampling of VGGNet, the information of some smaller objects are filtered out, resulting in the incomplete information of the feature map input into the regional proposal network [7, 8] In view of these shortcomings, an object detection algorithm based on dense network and multilayer information fusion is proposed. 2. Research on Object Detection Algorithm Based on Multilayer Information Fusion e detection flow of this algorithm is shown, which mainly includes four parts: (1) multilevel feature extraction; (2) multistage feature dimension specification and feature fusion; (3) region of interest extraction, classification, and regression; and (4) calculation of multitask loss. Suppose that a dense block has l layers in common, and each layer of network contains a nonlinear transformation Hi(x). e specific transformation is shown in (1)

Part 4 Multitask loss function
Experiment and Discussion
Findings
Contrast Experiment
Conclusion and Outlook
Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call