Small Target Object Detection with Transformer (STO-DETR) Algorithm Based on Swin Transformer

With the development of deep learning, the method based on Convolutional Neural Network (CNN) has made a series of major breakthroughs in the detection field. However, due to the shortcomings of low image resolution, small object size, inconspicuous feature information, and lack of small target data, the detection of small objects is still a challenging task in the field of computer vision and a hot research topic worldwide. To improve the accuracy and speed of small object detection, a new algorithm model of Detection with Transformer (DETR) small target object detection based on Swin Transformer is proposed in this paper. This paper proposes a new algorithm for small target object detection based on Swin Transformer, called Swin Transformer’s DETR, which is different from DETR which utilizes the CNN as the backbone network to extract the feature information. It extracts the features of the input image by the Swin Transformer backbone and reduces its size, and then reshapes the output feature map as C × H W, and a 1 × 1 convolution reduces the dimension, and then input to Transformer Encoder. Self-Attention is used to make global analysis on the feature graph, and Transformer Decoder is used to realize one-to-one correspondence between the target prediction frame and token. The output results of each layer of the Decoder are also predicted by the last Feed Forward Networks (FFN) shared by parameters, and the loss is calculated, thus realizing deep supervision. Finally, several potential directions of this research field in the future are provided, and the experimental results of this detection algorithm are compared.