| With the rapid development of information technology and the Internet,various media data have also grown massively,and a large amount of visual data has continuously emerged.In order to apply visual data more efficiently and quickly,many researchers began to study computer vision technology systematically.As one of the important basic tasks of computer vision,object detection has very important theoretical and applied value in many fields such as video surveillance,military observation,medical diagnosis and automatic driving,so it has gradually become a hot research problem in the field of computer vision and artificial intelligence.With the continuous research on this issue in academia and industry,the detection accuracy and speed of the object detection algorithm have been significantly improved,but there are still many challenges in the practical application: In real scenes,the collected images,videos,etc.often inevitably have a complex background.For example,some special backgrounds are very similar to the shape of the target to be inspected,and can easily be misjudged as a target to be inspected,which will undoubtedly increase the probability of the target to be wrongly identified;In real scenes,the distribution of objects is messy and the angle of observation is different.The target object may have different degrees of occlusion,which increases the possibility of missing detection;Due to the problem of shooting distance and Angle,the size of the target in the image also varies greatly,which has a certain influence on the accuracy of the location of the target area.These problems all increase the difficulty of detection and seriously affect the performance of detection.Aiming at the above challenges,this thesis studies the object detection algorithm in turn.The main work contents and innovation points are summarized as follows:(1)Aiming at the problem that the target is wrongly checked and missed due to the complex background interference and occlusion in the image,this thesis introduces the attention mechanism into the object detection,and forms the detection method of the target attention mechanism.By focusing on the target to be detected,the clutter and irrelevant targets are filtered out to reduce the complexity of the detection task;In addition,for the occlusion problem in the image,this thesis improves the NMS post-processing method that may cause overlapping occlusion targets to be missed.The Soft NMS method is used to replace the traditional NMS method for post-processing,so that the model can be more robust for detection of overlapping occlusion targets;Finally,in order to further reduce the calculation cost and improve the detection rate,this paper lightens the network architecture of the detection model,In addition,the ROI Pooling operation in this process is improved by referring to the ROI Align Pooling technology for the problem of inaccurate positioning caused by the deviation of candidate boxes caused by quantification.(2)Aiming at the problem of scale difference of target objects,this thesis improves the SSD multi-scale independent detection model which focuses on speed,fuses network features by feature pyramid method and fuses channel attention mechanism for object detection.The main idea of the improved model is as follows: firstly,feature extraction of the input image is carried out using convolutional neural network;Then,the feature pyramid method is used to fuse the feature of high convolutional layer with high semantics into the feature of low convolutional layer with high resolution layer by layer,so as to enhance the semantic information of the feature of each convolutional layer and obtain more expressive features.Then,the channel attention module is fused with the feature maps of different scales after fusion to model the correlation among the channels of the feature maps and further optimize the features to obtain more comprehensive expressive features,so as to enhance the network's ability to deal with multi-scale targets. |
PDF Full Text Request