Method Research On Object Detection In Real Scene

The object detection is an important basic subject in the field of Computer Vision.It is not only the basis of research topics such as behavior detection and image segmentation but also widely used in practical applications such as intelligent monitoring and smart driving.This thesis focuses on further research on the general algorithms in multi-target detection tasks.The main contents are as follows:First,it is determined that the algorithms in this thesis are based on Convolution Neural Network(CNN)by comparing the traditional object detection algorithm and deep learning-based object detection algorithm.In addition,typical algorithms and basic structures of CNN are briefly introduced and discussed,which lays a foundation for the latter algorithm.Second,the proposed algorithm in the thesis based on the Single Shot MultiBox Detector(SSD)algorithm and uses the strategy of feature reuse.A modified convolution feature pyramid is proposed to construct high-resolution feature maps with semantic information which can be extracted in high-level feature maps.The result shows a significant improvement in mAP.For better performance,the impact of using different upsampling methods to construct feature pyramid is discussed and a fast and accuracy upsampling method is proposed.Third,a novel attention mechanism based on pixel position is proposed.The novel attention algorithm uses the high-level feature information to construct a map to weight the low-level feature maps,which force the network focus on feature extraction on the target region.The experiments show that this algorithm can improve detection accuracy slightly.Fourth,in order to further improve the classification accuracy,a lightweight customized structure is added to the classification branch,which can improve the effectiveness of the models.At the same time,the loss function is modified due to the problem of imbalanced training data to improve the detection result of the less classes.