针对新学习策略的弱小目标检测方法

Dim small target detection method based onnew learning strategy

摘要:

基于深度卷积神经网络的目标检测在应用中展现出了良好的性能,然而,将其应用于弱小目标检测上依然性能欠佳。本文提出一种有效的弱小目标检测方法,使用改进的特征提取方法,利用尺度匹配策略选取合适的尺度进行小目标检测。同时在神经网络中设计自适应的融合模块,通过融合特征与接收域以增强目标环境特征。提出的方法在检测速度和精度上都具备良好的性能。有效解决了一般的框策略无法准确获取小目标的问题,新的策略使用自适应参数确定检测框。实验结果表明,提出的方法在视频数据中能够有效检测弱小目标,优于其它先进的目标检测方法。

Abstract:

Target detection based on deep convolution neural network shows good performance in application. However, its performance in weak and small target detection is still poor. In this paper, an effective weak and small target detection method is proposed. The improved feature extraction method and scale matching strategy are used to select the appropriate scale for small target detection. At the same time, an adaptive fusion module is designed in the neural network to enhance the characteristics of the target environment by fusing the features and the receiving domain. The proposed method has good performance in detection speed and accuracy. It effectively solves the problem that the general box strategy can not accurately obtain small targets. The new strategy uses adaptive parameters to determine the detection box. Experimental results show that the proposed method can effectively detect weak and small targets in video data, and is superior to other advanced target detection methods.