Deformable Perceptual Attention Mechanism And Pyramid Features For Image Object Detection

Object detection is a fundamental branch of computer vision that provides valuable information for complex visual processing tasks such as scene understanding,image description and object tracking.The task of object detection is to judge the category of objects in the image and determine the position in the image,which is one of the most basic and challenging problems in the field of computer vision.In recent years,deep learning technology has made a breakthrough in the field of object detection by virtue of its strong feature representation ability.Therefore,the object detection algorithm based on deep learning has become the mainstream method.How to extract the objects features with rich semantic information and design a reasonable and efficient object detection network structure,which is the core method to improve the accuracy of the object detection algorithm.This paper focuses on the following aspects of the core problem.1.A object detection network based on depth adaptive pyramid features and hierarchical supervision is proposed.Aiming at the gap problem of feature semantic information in multi-scale feature fusion,an adaptive feature fusion module and a hierarchical monitoring mechanism are designed on the basis of feature pyramid network FPN.On the one hand,adaptive learning channel weights are used to combine pyramid scale features,on the other hand,hierarchical supervision mechanism is used to implement the same monitoring signal for multi-scale features,which reduces the semantic gap between different scale features.The average accuracy of the PSAVAL VOC data set and Vis Drone2019 data set was 81.45% and 20.18% respectively,which was 1.76% and 0.66% higher than the two-stage object detection algorithm represented by FPN,and 2.87% and 1.69% higher than the single-stage object detection algorithm represented by Retina Net,respectively.2.A object detection algorithm based on deformable perceptual attention mechanism is proposed.Aiming at the problem of image background interference object detection,an attention mechanism is designed to change the attentional perception region with the change of the shape of the object,which can effectively extract the detail information of the specific region in the image.By integrating the channel information and spatial information of the features,the attention mechanism can obtain the attention perception features,strengthen the semantic information of the salient features and weaken the background interference information.When combined with Faster R-CNN algorithm,thePSAVAL VOC data set was improved by 1.24% and 1.53% respectively,compared withthe channel attention mechanism and the spatial attention mechanism.Combined with thechapter 3 algorithm,the precision on the Vis Drone2019 dataset is improved by 0.04%.Therefore,the deformable perceptual attention mechanism can effectively assist the network to extract important information in a specific area,improve the recognition of attentional perceptual features,and improve the detection accuracy.3.A object detection network based on deep dilated convolution and hierarchical supervision is proposed.In order to improve the ability of multi-scale feature expression from the perspective of network design,the receptive field expansion module is designed based on FPN of feature pyramid network.The receptive field expansion module is formed by the combination of multi-branch dilated convolution and standard convolution,so that the acquired features can have both global information and local details.The accuracy of the PSAVAL VOC data set was gradually improved by adding the sensing field development module on each layer,up to 1.04%.Compared with the FPN algorithm,it is improved by 0.27% on the Vis Drone2019 dataset and reduced by 0.39% compared with the algorithm in chapter 3.