Research On Object Detection Algorithm Based On Dual-path Convolutional Neural Network

In the field of computer vision,object detection algorithm is one of the basic tasks.It has a wide range of application value in the fields of automatic driving,industrial detection,text detection,pedestrian detection,face detection,garbage classification,remote sensing image detection,medical image detection and so on.The object detection algorithm has achieved leapfrog development in recent years,and the object detection algorithm has become more and more powerful.With strong motivation,the object detection algorithm based on deep learning has room for further research and improvement,and there are many problems that need to be optimized,such as how to design the structure of the convolutional neural network to extract object information that is more suitable for object detection tasks,and large-capacity data.The problem of sample imbalance in the set sample,how to solve the multi-scale detection problem of the object detection algorithm,how to perform effective feature fusion on the extracted convolutional neural network features to improve the detection accuracy,and how to improve the modeling of the detection frame.Therefore,we study the existing theoretical algorithms of object detection,and propose a dual-channel convolutional neural network suitable for object detection in response to some above problems,and also integrate some more advanced technologies and ideas to continuously optimize.The detection effect of the model was tested,and the training and testing were carried out on the public data set.A control experiment was designed to prove the rationality of the network structure,and the training and testing process were experimentally analyzed.The main work of the thesis are as follows:(1)In order to extract features suitable for the object detection task and solve the problem of multi-scale object detection,we propose a new dual-channel multi-scale object detection paradigm,and on this basis,we design a single-stage general-purpose anchor frame-based Object detection algorithm,named Dual Path Single Shot Detector(DPSSD).Dual channels can ensure that shallow features can be more easily used in detection,that is,residual channel and channel combination channel,to improve detection accuracy.Our improved dual-path basic network is more suitable for multi-scale object detection tasks.Together with the feature fusion module,it constitutes a multi-scale feature learning paradigm called dual-path feature pyramid.We use PASCAL VOC dataset and COCO dataset respectively.Models of 320 pixels and 512 pixels are trained,and the structures in the network are validated experimentally for inference.The experimental results show that our algorithm has an advantage in single-stage object detection algorithms related to SSD(Single Shot Detector),and has achieved an advanced level in average accuracy.(2)Variations in real-world images,including illumination,pose,deformation,background clutter,occlusion,blur,resolution,noise,and camera deformation,are a key challenge for object detection.Target objects usually coexist with other objects and environments,and background information,e.g.object relationships and global scene statistics,helps object detection and recognition,especially for small objects,occluded objects,and objects with poor image quality.In order to solve these problems and improve the robustness of features,we design different deep and shallow feature fusion modules for research.The features extracted by the convolutional neural network are fused and then predicted,which will significantly improve the robustness of the features.Therefore,we have conducted in-depth research on the fusion method between the features of different scales in the feature pyramid of the convolutional neural network.By reading the literature,summarizing the effective structures used in the relevant literature,we designed 12 feature fusion structures and conducted a control experiment to select the most suitable feature fusion structure for two-pass convolutional neural networks.(3)Some commonly used object detection algorithms need to preset anchor frame parameters according to experience,including the length-width ratio,number and other parameters of the anchor frame.In order to simplify this process,we no longer use the anchor frame to generate the suggested area,but use the real The method of box segmentation design suggested regions,read the latest research literature,and introduced a region modeling method based on pre-selected points,preset some regions that may contain objects,and then divided these regions into positive samples and negative samples,input into the neural network for training,and obtain a theoretical model with better detection effect.We introduced the parameter values used in the entire training phase of the model in detail,and visualized the loss function value and the change of each category confidence value during the testing process.