Real-Time Object Detection Based On Spatial Information Recovery And Semantic Information Enhancement

Object detection aims to locate and classify objects in images.It is an important foundation for the computer vision system to perform image understanding and make system decision.Object detection is widely used in fields such as autonomous driving,intelligent monitoring and human-machine interaction.It is a challenging task to achieve object detection with real-time speed and high detection accuracy.The anchor-free object detection algorithms based on key-points is an important object detection technology that has emerged in recent years.When using compact deep backbones,such algorithms can meet real-time speed but have some problems.The main problems include the loss of spatial detail information which leads to inaccurate location and weak feature semantic information which leads to the low confidence scores even misclassification.To tackle the above-mentioned location problem and classification problem,this thesis focuses on anchor-free object detection algorithms based on key-points and proposes two algorithms.The main works,innovations and contributions are as follows:(1)A hierarchical spatial information recovery network(HSIRNet)is proposed in this thesis.HSIRNet designs an adjacent layer information strength module in the encoder of network.It makes full use of the feature to improve the spatial detail information for the adjacent high-semantic and small-resolution feature and enhance the semantic information for the adjacent low-semantic and large-resolution feature.HSIRNet also designs a residual attentional feature fusion module in the decoder of network.It hierarchically fuses the low-level and high-level features better and recovers the loss of spatial detail information.The proposed method significantly alleviates the problem of inaccurate location and spatial detail information lost,which is because the traditional anchor-free object detection algorithms based on key-points directly upsample small-resolution features and use them for detection.HSIRNet achieves a better trade-off between the detection speed and the detection accuracy.However,it still has the problem that the detected objects have low classification confidence scores and even are misclassified.(2)In view of the fact that HSIRNet has poor classification performance,a semantic information enhancement network(SIENet)is proposed.SIENet designs a residual attention upsampling module in the decoder of network.It effectively uses the correlation of the local features to enhance the semantic information of features.SIENet also designs a channel semantic information enhancement module between the encoder and the decoder.It converts information of channels to spatial dimension,constructs multi-scale features and increases the receptive field of features.While ensuring real-time speed,SIENet can significantly alleviate the problem of low object classification confidence scores and misclassification due to weak semantic information of features.In conclusion,this thesis proposes the real-time object detection algorithm based on spatial information recovery and semantic information enhancement.It effectively alleviates the problem of spatial detail information lost and weak semantic information which exists in the anchor-free object detection algorithms based on key-points.The proposed algorithm improves the detection accuracy while ensuring real-time speed.