Small object detector based on hybrid transformer architecture

Deep learning has developed rapidly in recent years. Convolutional neural networks perform well in a series of visual downstream tasks, and perform fast while achieving excellent accuracy. However, small object detection is still one of the difficulties in target detection tasks. The receptive field of the convolutional neural network is limited by the size of the convolution kernel, and it is difficult to capture the global dependency, which in turn affects the detection accuracy of small objects. Therefore, the researchers began to introduce the Transformer structure, using the advantages of selfattention in global modeling capabilities to improve small object detection capabilities. However, the secondary computational complexity of the self-attention mechanism greatly increases the computational consumption of the model. Subsequent research began to design a hybrid architecture model of CNN and Transformer to combine the advantages of both. It has the long-distance dependence capture capability of Transformer while having the execution speed close to CNN, and improves the detection accuracy of small targets. We propose a new target detector YOLOFormer-SD, which combines the Transformer structure in the YOLO series model. In order to balance performance and speed, we use different token mixers in the backbone and propose Residual-MLP to improve the model 's ability to capture local dependencies. The experimental results on the public data set Visdrone show that YOLOFormer-SD outperforms the existing YOLO series models with similar parameters in performance, and the delay is not greatly improved, showing its competitiveness in the field of small object detection.