Research on small target detection technology based on slicing aided inference

DOI:	10.11809/bqzbgcxb2024.02.035
Keywords:	target recognition; small object detection; feature fusion; sliced inference; loss function
Abstract:	Aiming at the problems of low recognition accuracy and high detection rate of small targets in high resolution images, this paper optimizes three aspects from the perspective of model training and prediction, including feature fusion, prediction block position loss function and detection network, and proposes an improved YOLOv5 algorithm based on slice reasoning.Firstly, the CAM module was added to the feature fusion network, and the scale of the feature receptive field was increased by expanding convolution, which enhanced the contextual learning effect of the small target and its adjacent pixels.Secondly, the Focal_EIOU loss function was used to replace the original CIOU loss function, which minimized the difference of width and height between the prediction frame and the real frame, paid more attention to the effective target prediction results, and improved the positioning accuracy of the prediction frame and the convergence rate of the loss function.Finally, the SAHI algorithm was added to the detection network, and the slice idea was used to magnify local features and predict the slice results respectively, which improved the detection effect of local features and reduced the omission rate of small targets.Compared with the original YOLOv5 algorithm, the bounding box positioning loss is significantly reduced and the convergence is faster. The recognition accuracy of small targets is improved by 4.4%, and the detection rate of small targets is twice of the original, which can be effectively applied to the small target detection task in high resolution images.
Published:	2024-02-28
PDF