基于改进F3Net网络的显著性目标检测

Salient object detection based on improved F3Net model

针对F3Net网络中因特征抽取导致的空间分辨率损失问题,采用一种锥体卷积和感受野模块相结合的方法,有效减少空间分辨率的损失.该方法中锥体卷积能够以不同深度或大小的卷积核并行处理输入图像,从而考虑不同级别空间的上下文依赖关系.此外,编码器提取的最高层特征在通过感受野模块后可以捕获更多的目标细节.实验结果表明,改进后的F3Net网络生成的图像质量明显提高,检测性能得到改善.

Aiming at the problem of spatial resolution loss caused by feature extraction in F3Net network,a method combining pyramidal convolution and receptive field module is adopted to effectively reduce the loss of spatial resolution.Pyramidal convolution can process the input images in parallel with convolution kernels of different depths and sizes,so as to consider the context dependence of different levels of space.In addition,the highest level features extracted by the encoder can capturemore target details after passing through the receptive field module.Experimental results show that the image quality and detection performance of the improved F3Net network are improved obviously.