Research And Application On One-shot Object Detection Based On Attention Mechanism

In recent years,with the increase in the amount of data and the improvement of computer computing capabilities,deep learning-based object detection algorithms have made major breakthroughs.However,the performance of a deep learning model depends on a large amount of training data.For an object detection task with a small sample size of training data,a network with a large number of parameters is used to solve it,which is prone to overfitting.On the other hand,object detection does not have a good detection effect for the categories that are not included in the training sample with a small amount of data.Aiming at these two problems,this thesis proposes a one-shot object detection optimization algorithm.The main research contents and innovations are as follows:1)The attention mechanism is applied in the object detection model.For the feature pyramid modules currently used in many networks,this paper proposes a feature fusion module based on the channel attention mechanism to better fuse high-level and low-level feature.At the same time,when designing a new feature pyramid,a new feature enhancement module is designed for the lower feature layers,considering the weak semantic information of low-level features.In actual use,we embed these two modules into an object detection model.In this paper,a large number of comparative experiments are performed on the public data sets,VOC and COCO,to verify the effectiveness of the feature fusion module and feature enhancement module.2)This paper proposes a one-shot object detection model.Given a template image containing only an object,the model can find the position of other objects belonging to the same category in the test picture,so that the model has a good generalization on the test set of unknown category.At the same time,the newly designed feature fusion module,FFM,is fused into this model and detected on feature maps of multiple scales.Experimental results show that the model,proposed in this paper,significantly improves the accuracy of the one-shot object detection algorithm.3)Aiming at the problem that the one-shot object detection network,will judge the background as an object,a false positive case filtering model is proposed.The post-processing method will further filter the output results,and effectively reduces the number of false positive cases output by the model and improves model performance.