Research On Object Detection Algorithm Based On Feature Learning

At present,object detection has great application requirements in intelligent systems(such as unmanned driving,robotics,monitoring and monitoring,etc.),image content classification and retrieval,and auxiliary medical treatment(such as medical image lesion extraction and analysis).However,the traditional detection algorithm is difficult to break through in performance due to the limitations of feature design.The deep learning was first applied to the field of computer vision,and with the obvious advantages,it has been more and more loved by researchers from all walks of life.In a short time,various algorithms and network models for target detection emerge in an endless stream.Improve target detection performance.This paper takes the object detection of static images as the research goal.Through research and analysis of various object detection network based on deep learning that have been proposed,the improvement and innovation of the deficiencies are made.The main research contents and innovations of this paper are as follows:1)Through the research on the principle and performance of the proposed object detection network based on deep learning,this paper proposes a small object detection algorithm based on improved Faster R-CNN to improve the detection effect of small objects.This paper first analyzes the reasons for the poor detection of small objects in the Faster R-CNN network,and adopts a joint mechanism to improve its detection performance.This method combines the generative adversarial networks model in the Faster R-CNN framework,and uses the generation network to map the low-resolution small object features to high-resolution features with stronger expression,use the network itself for good detection of medium and large objects to meet the requirements of improving overall detection performance.In this paper,the improved network is tested on the Logo dataset.The experimental data shows that the improved network is better than the Faster R-CNN in terms of detection performance,indicating that the improved method plays a role.2)Through research,it is found that a network using only simple high-level features has been difficult to achieve breakthroughs in object detection performance.A network that fuses multiple features can provide more feature information when selecting object prediction boxes.Therefore,this paper proposes a new detection network that combines multi-scale features and global contextual features.By using the rolling convolution method in Faster R-CNN,multi-scale features are obtained while ensuring the conciseness of the network structure.And a Recurrent Neural Network of multiple gated recursive units(GRUs)is used at the top of the convolutional layer to highlight the useful global contextual information in the feature map,further improving the object detection performance.Comparative experiments on the benchmark datasets PASCAL VOC 2007 and PASCAL VOC 2012 show the superiority of the proposed method.