针对小目标(像素占比小于0.02)检测存在的目标特征容易丢失、分辨率低的问题,提出了一种基于改进YOLOv3(You only look once)卷积神经网络的检测方法。首先,对数据集中的小目标进行复制变换增强,以提升训练过程中网络对小目标的注意力。其次,针对浅层视觉信息与深层语义信息的尺度融合,提出了跨尺度检测层的网络结构,提高了网络对小目标的适应能力。最后,针对高分辨率图像的检测效果,提出了深度和广度结合的残差块组传递结构,丰富了深层特征图的感受野。实验结果表明,相比YOLOv3网络,改进跨级尺度预测层的网络检测小目标的精确率提升了1.9个百分点,召回率提升了5.9个百分点;优化感受野的网络检测小目标的精确率提升了31.6个百分点,召回率提升了46.4个百分点。

Aiming at the problem of small target (pixel ratio less than 0.02) detection that the target features are easily lost and the resolution is low, a detection method based on improved YOLOv3 (You only look once) convolutional neural network is proposed in this paper. First, the small targets in the data set are copied and transformed to enhance the network''s attention to the small targets during the training process. Second, for the scale fusion of shallow visual information and deep semantic information, a cross-scale detection layer network structure is proposed, which improves the network''s adaptability to small targets. Finally, for the detection effect of high-resolution images, a residual block transfer structure combining depth and breadth is proposed, which enriches the receptive field of deep feature maps. Experimental results show that compared with the YOLOv3 network, the precision rate of the network detection of small targets with the improved cross-scale prediction layer increased by 1.9 percentage points, and the recall rate increased by 5.9 percentage points. The precision rate of the network detection of small targets with the optimized receptive fields increased 31.6 percentage points, the recall rate increased by 46.4 percentage points.