| Visual object detection is one of the areas of research in which artificial intelligence techniques have received attention.In recent years,deep learningbased object detection techniques have made far-reaching progress with the proliferation of data,the proposal of excellent detection frameworks,and the development of computational power,but they are also limited by this.There are two broad bottlenecks affecting algorithm performance: first,as detection tasks become more complex,networks are designed to be deeper and deeper,and many high-performance object detection models are unable to be deployed on less powerful devices or have lower detection rates due to the complexity of the algorithms,resulting in a significant reduction in their engineering application value.Secondly,existing object detection algorithms are not ideal for small-scale objects,which occupy fewer pixels,and if the feature extraction network is not fine-grained enough to extract the features,the detection results will be missed.In this thesis,based on the existing classical object detection algorithms,we explore a real-time object detection algorithm that balances detection accuracy and detection rate by designing a lightweight network model to address the above problems.The research in this thesis is as follows:1.In this thesis,an in-depth study and summary of the Anchor-based algorithm is conducted to design a lightweight object detection algorithm based on model compression for YOLO v3.First,a clustering operation is done on the real annotated samples in the dataset to obtain a total of 9 Anchor boxes at 3scales.In the network design phase,the idea of ECANet is introduced on the lightweight feature extraction network to weight the features of the object and filter invalid information such as background.The multi-scale feature fusion stage uses three scales of feature maps to first perform top-down feature fusion,so that the deep semantic information is fused to the shallow layer.A further bottom-up feature fusion is performed so that the more detailed feature maps of the lower layers are convolved and then fused with those of the higher layers.The CIo U Loss is introduced to alleviate the phenomenon that the loss of 0 fails to gradient back when there is no overlap between the prediction box and the real box.Finally,the model is lightened by model compression techniques.2.To meet the demand of real-time detection of edge devices,this thesis analyzes the Anchor-free algorithm implementation and proposes a real-time target detection algorithm based on TTFNet with enhanced feature fusion.The backbone network uses Mobile Net,a lightweight network,as the feature extractor,and focuses important information with high weights and filters non-critical information with low weights through an efficient attention mechanism before feature fusion,so that the model focuses more on foreground semantic information.In the feature fusion stage,the feature maps extracted from the backbone network are optimized and connected across layers in the upsampling stage by multiple one-dimensional convolutions in asymmetric convolution to enhance the semantic information of the feature maps.For the problem of degradation of the prediction frame position regression loss of the original method,it is mitigated by introducing DIo U Loss.Finally,ablation experiments are designed and the effectiveness of the algorithm is demonstrated.In summary,the proposed algorithm is fully experimented on publicly available datasets and tested on mobile devices.The experimental results show that the proposed algorithm achieves the detection rate of real-time detection while maintaining the original detection accuracy,which is conducive to the low computing power and low memory requirements of mobile devices.At the same time,it has better robustness than the original method. |
