Research On Object Detection Algorithm Based On Multi-scale Feature In Complex Scene

In the age of artificial intelligence,traditional object detection methods have been unable to cope with a large amount of image and video data in terms of accuracy and realtime performance.With the development of artificial intelligence,object detection technology based on deep learning arises at the historic moment,showing obvious advantages in many general and specific detection tasks.However,many detection results such as false detection and missed detection in complex scenes also reveal some limitations of this method.By analyzing the common problems of object detection in complex scenes,we assist the object detection framework to process single frame image quickly and accurately from the perspective of multi-scale features,so as to further improve the performance of the detector.The main contents in this paper are as follows:(1)An image pyramid object detection algorithm based on dual attention mechanism is proposed.Based on the backbone network SSD(Single Shot Multibox Detector),the image pyramid auxiliary network and dual attention mechanism are designed.The auxiliary network built by the image pyramid only performs shallow convolution operations on the image,so the multi-scale features extracted by the network retain rich scene information,thereby improving the spatial features lost in the deep layers of the backbone network.Meanwhile,the proposed algorithm can adaptively filter the redundant features of backbone network and auxiliary network with the help of dual attention mechanism.Experiments show that the detection accuracy and inference time of the algorithm are 32.6% AP and 23 ms,respectively.Compared with other algorithms,the algorithm achieves a better balance in complex scenes.(2)An object detection algorithm based on multi-scale feature interaction is proposed.Combined with the key contents expressed by different scale features,the algorithm builds a multi-scale interaction module,which completes the gradual interaction among features from shallow to deep.Moreover,the feature-aware enhancement module is designed to further enhance the feature representation of small objects in complex scenes by amplify the differences between the interactive features and the original backbone features,so as to achieve a more accurate detection result.Experimental results of multiple data sets show that the proposed algorithm is effective in detecting small objects.(3)A multi-scale object detection algorithm based on dynamic label assignment is proposed.The algorithm modifies the Io U(Intersection over Union)between regression box and ground truth box with the help of dynamic weight to achieve the optimal allocation of positive and negative samples.With the support of dynamic label assignment,the object detection framework is constructed by auxiliary correction mechanism and feature guidance module.The auxiliary correction mechanism can supplement the information of the auxiliary network to the backbone network,and guide the subsequent feature extraction of the backbone network.The feature guidance module realizes the effective interaction among multi-scale features under the guidance of feedback learning.Experiments show that the algorithm achieves an effective allocation of positive and negative samples in complex scenes while obtaining 35.2% AP and 82.0% AP on MS COCO and Pascal VOC datasets,respectively.