A Weak and Small Target Detection Algorithm Based on the View of UAV

The target has weak feature and small size in the view of UAV(Unmanned aerial vehicle), so This paper proposes an algorithm of weak and tiny target detection in the view of UAV. Firstly, The method use the contrast limited adaptive histogram equalization to enhance the image, improve the weak target features, weaken the background, and make the model more suitable for different scenes. Secondly, an parameterless SimAm is introduced to assign weights to different pixels of different feature maps to enhance features. Then, The method use Pixel Recombination downsampling module to avoids the problem of fine-grained information loss caused by convolution operation and pooling operation, and ensure the detection performance of low resolution image and small target. At the same time, the bounding box regression loss with dynamic focusing mechanism is introduced to reduces the competitiveness of high-quality anchor and the harmful gradient generated by low-quality examples, so that the loss function can focus on ordinary quality anchor and improve the overall performance of the detector. Finally, the model is pruned to ensure the lightweight of the model and the real-time detection. The effectiveness of the improved algorithm was verified on the dataset VisDronr2019. The experimental results show that the parameter number of the improved algorithm is reduced to 81%, [email protected] is increased by 7.4%, and [email protected]:0.95 is increased by 2.9%. This shows that the improved method can effectively improve the detection accuracy of weak and small targets.