Research On UAV Small Target Detection Algorithm Based On Improved Convolutional Neural Network

Uavs(Unmanned Aerial Vehicle)are widely used in various fields because of their wide field of vision,good mobility and low cost.However,UAVs are in high flight,and their shooting targets are small and their features are fuzzy,which leads to poor accuracy of target detection by UAVs.In order to solve the above problems,VGG16(Visual Geometry Group Network)and YOLOv5s(You Only Look Once v5s)are improved in this paper and applied to UAV target detection.Firstly,VGG16 is adopted as the identification network of target categories,the Re LU activation function in VGG16 is replaced by Sigmoid function to increase the network fitting ability,and GAP(Global Average Pooling)is used to replace Flatten to reduce the number of network parameters.At the same time,the Adam(Adaptive Moment Estimation)step length optimizer is used for model training optimization.The training and anti-noise comparison experiments between VGG16 and the improved VGG16 were carried out on the self-made data set.The experimental results show that under the same training times,VGG16 underfit occurs,and the improved VGG16 accuracy converges around 1,and the Loss tends to 0.05.In addition,when the noise content is 50%,the accuracy of the improved VGG16 is about 33.21% higher than that of VGG16.Secondly,in order to effectively reduce the cost of UAV target detection system,lightweight improvement is carried out on the basis of YOLOv5 s.These include the use of deep separable convolution instead of traditional convolution to reduce the number of model parameters.The Dropblock to improve the noise resistance of the model is added.The bilinear interpolation up-sampling method is applied to enhance the model recognition capability.The attention-mechanism module SENet(Squeest-and-Excitation Networks)is used to improve the small target detection capability.The improved YOLOv5 s reduces the memory by 3.6M and achieves an 83% accuracy on homemade data sets.Thirdly,on the basis of the improved YOLOv5 s,the SENet attention mechanism is upgraded to the Inception-V4 with multi-scale convolution,and this module is combined with the improved VGG16.IVGGNBI(Improved Visual Geometry Group Network Based on Inception)is proposed.The upgraded YOLOv5 s is trained on COCO128 data set,and the trained model is used to identify the pictures in the self-made data set one by one and crop the pictures with the anchor frame as the boundary.This paper gathers the trimmed images into data sets and uses Ada BN(Adaptive Batch Normalization)algorithm on the basis of IVGGNBI to keep training and anti-noise experiments on data sets.Experimental results show that the accuracy of the upgraded YOLOv5 s algorithm combined with the IVGGNBI network model based on the Ada BN algorithm reaches 97.8%.Finally,ablation experiments were conducted on the improvements of VGG16 and YOLOv5 s algorithms,and the results of ablation experiments verified the rationality of various improvements of VGG16 and YOLOv5 s.In addition,the experimental results show that when the image contains 70% noise,the model recognition accuracy of the upgraded YOLOv5 s and the Adab N-based IVGGNBI network can still reach about 90%.