Research On Small Object Detection Algorithm Based On Deep Learning

Object detection is an important research direction in computer vision,and it is of great significance in production applications and academic research.Object detection has developed rapidly in recent years.especially after the FPN(Feature PyramidNetworks)structure is widely recognized and adopted,the accuracy of object detection has been improved greatly.Efficient Det is a leader in the target detector based on the FPN structure.The detector has the characteristics of low parameters and achieves very high accuracy..But it was found that the fusion speed of high-level feature information and low-level feature information during the training process of Efficient Det is not fast enough after many experiments and theoretical deductions,and it relies on excellent hardware during training,especially,It depends on the multi-layer Bi FPN repeat block superposition and the increase of the input size and the number of channels to facilitate the completion of multi-scale feature fusion.In order to achieve faster convergence speed and higher accuracy,this thesis proposes WavesNet based on Efficient Det which mainly improves the characteristic network structure.First,seven channels are added on the basis of Bi FPN to make the features of each level merge more quickly and effectively,and then the Com C extraction method is added in the initial feature extraction stage of the feature network,in addition,the fast normalization formula for feature fusion proposed by Efficient Det is improved to ensure the stability of the feature value which is called stable fast normalization formula.Comparative experiments and ablation experiments were done,in order to verify the robustness and superiority of the network,the results show that WavesNet's convergence speed is about 30% higher than Efficient Det,and its accuracy is 0.5% higher than Efficient Det on COCO dataset.Experiments show that the WavesNet object detector has faster convergence speed and higher accuracy,especially has higher detection accuracy on small object.