基于轻量型卷积神经网络的海面红外显著性目标检测方法

Infrared salient object detection of sea background based on lightweight CNN

为提高红外舰船图像显著性检测精度,同时降低参数量,提出一种轻量型红外舰船显著性检测模型。该模型针对红外图像缺乏颜色、纹理等细节特征的特点,从以下三个方面进行轻量化设计:在骨干网络设计方面,将视觉几何组网络(visual geometry group, VGG)各层的通道数减少一半作为骨干网络,以减少冗余的特征;为了进一步减少模型参数量,在前两个低层卷积模块中引入一种轻量型的线性瓶颈模块(linear bottleneck, LB)替换传统卷积模块;提出一种新的提取全局特征能力更强的轻量型的高层线性瓶颈模块(high-level linear bottleneck, HLLB)替换后3个高层传统卷积模块,并且使用自适应平均池化提取高层特征作为全局特征以得到更丰富的上下文信息。针对红外数据集缺少的问题,构建一个红外舰船数据集IRShip,包括1002幅图像。试验结果表明:该算法能够有效实现红外舰船目标的显著性检测,并且通过与其他7种常用的显著性检测模型对比,本研究提出的模型可以在大幅减少参数量的情况下有效提升红外舰船显著性目标检测的性能。

In order to improve the saliency detection accuracy of infrared(IR) ship images and reduce the amount of parameters, a lightweight IR ship salient object detection model was proposed. By analyzing the characteristic of IR images, the lightweight model from the following three aspects was designed. In terms of backbone network, to reduce redundancy features, we applied VGG16 with half channel numbers as the backbone, which was obtained by reducing the number of channels in each layer by half. To further reduce the amount of parameters, we introduced the linear bottleneck, which was an efficient lightweight module to replace the traditional convolution module in the first two low-level convolution modules. It was beneficial to improve the efficiency. To improve the ability of capturing global features, we developed a new lightweight module named high-level linear bottleneck(HLLB) to replace the last three high-level traditional convolution modules, and adaptive average pooling was applied to obtain more high-level context information. Due to the problem of lacking IR dataset, we constructed an IR ship dataset(IR Ship), including 1002 images. The experimental results showed that the proposed method could effectively detect the salient IR ship targets. Comparison results with other 7 commonly used saliency detection demonstrated that the proposed method could achieve better detection results with much fewer parameters than other methods.