Small Object Detection Model Based On Deep Learning And Its Lightweight Research

The object detection algorithm has been consistently enhanced and applied to diverse scenarios of life in the era of deep learning and the detection accuracy has been continuously improved.The majority of current object detection algorithms,on the other hand,are primarily focused on detecting medium and large objects,resulting in low detection accuracy for small objects.Furthermore,the neural network model becomes more complicated as the real-world issues to be solved become more varied,and the number of parameters and computations increases.These issues directly contribute to the model’s difficulties in being implemented on devices with limited computing power and the slow detection speed in the CPU environment.This thesis focuses on related research on the two topics mentioned above and proposes the small object detection model AFSDet,which is based on attention mechanisms and feature fusion and addresses the problem of indistinguishable small object and image background,as well as considerable feature loss during feature extraction.AFSDet makes use of the attention mechanism to improve the backbone network’s feature extraction capabilities and combine the features of multiple layers by skip connections.Simultaneously,a lightweight research on AFSDet is carried out to address the problem that the neural network model is difficult to apply to devices with limited computing power due to the huge number of parameters and computation.From the standpoint of the backbone network and detection heads,this research introduces LFSDet,a lightweight small object detection model that compresses the parameters and computation of AFSDet.This thesis examines the properties of various lightweight backbone networks,chooses Shuffle Net V2 as the model’s backbone network,and presents the Light-Head lightweight detection head.This research conducts considerable comparative experiments on the general dataset Pascal VOC and the remote sensing dataset NWPU VHR-10 in order to verify the model’s effect.The small object detection model AFSDet proposed in this research improves the detection accuracy indicators by 1.4% to 19.2% when compared to the baseline model Center Net,with the small object detection accuracy improving the most,indicating the usefulness of AFSDet for small object detection.Furthermore,the impact of the model compression strategy used in this thesis is extremely clear.The number of parameters in the lightweight small object detection model LFSDet is decreased by 78.1 percent,the amount of floating point operations is lowered by 83.7percent,and the model size is compressed by 78.9 percent when compared to AFSDet.Ablation experiments are also used to verify the rationale and correctness of the AFSDet submodule and compressing model from various perspectives.