AFFPN: Attention Fusion Feature Pyramid Network for Small Infrared Target Detection

Abstract

The detection of small infrared targets lacking texture and shape information in the presence of complex background clutter is a challenge that has attracted considerable research attention in recent years. Typical deep learning-based target detection methods are designed with deeper network structures, which may lose targets in the deeper layers and cannot directly be used for small infrared target detection. Therefore, we designed the attention fusion feature pyramid network (AFFPN) specifically for small infrared target detection. Specifically, it consists of feature extraction and feature fusion modules. In the feature extraction stage, the global contextual prior information of small targets is first considered in the deep layer of the network using the atrous spatial pyramid pooling module. Subsequently, the spatial location and semantic information features of small infrared targets in the shallow and deep layers are adaptively enhanced by the designed attention fusion module to improve the feature representation capability of the network for targets. Finally, high-performance detection is achieved through the multilayer feature fusion mechanism. Moreover, we performed a comprehensive ablation study to evaluate the effectiveness of each component. The results demonstrate that the proposed method performs better than state-of-the-art methods on a publicly available dataset. Furthermore, AFFPN was deployed on an NVIDIA Jetson AGX Xavier development board and achieved real-time target detection, further advancing practical research and applications in the field of unmanned aerial vehicle infrared search and tracking.

LW-IRSTNet: Lightweight Infrared Small Target Segmentation Network and Application Deployment

“…Although IR small targets occupy very small pixels in the image, there are also changes in different scales. To enable the network to learn the characteristics of IR small targets of different scales as much as possible, LSPM [25] and AFFPN [27]construct multi-scale feature maps on the high-level feature maps using atrous convolution and adaptive global average pooling. A bidirectional semantic segmentation network is proposed, and a feature fusion module is introduced to achieve a balance between speed and segmentation performance.…”

“…[50], [22], [25]. (c1) -(c7) show different contextual feature fusion modules, where (c3) -(c7) are derived from [19], [22], [27], [51], [52], respectively.…”

“…Huang et al [25] used VGG [26] as the backbone to design a local similarity pyramid module (LSPM) to fit the multiscale features of small infrared targets. Zuo et al [27] designed a multiscale feature fusion pyramid module (AFFPN) with ResNet as the backbone to deal with the loss of targets in the deep network. Wang et al [28] used ResNet as the backbone, used region proposal network (RPN) to extract candidate targets, used FCN [29] network to extract feature maps, and used Transformer [30] to determine candidate targets.…”

LW-IRSTNet: Lightweight Infrared Small Target Segmentation Network

“…3) Multi-Scale Feature Fusion Strategy: Although small infrared targets occupy very small pixels in the image, there are also changes in different scales. To enable the network to learn the characteristics of small infrared targets of different scales as much as possible, inspired by the ASPP module in Deeplab network, AGPCNet [22], LSPM [25], and AFFPN [27], construct multi-scale feature maps on the high-level feature maps using hole convolution, adaptive global average pooling, etc., and then splice them through the concat function, and finally through 1×1 convolution for feature fusion.…”

“…Huang et al [25] used VGG [26] as the backbone to design a local similarity pyramid module (LSPM) to fit the multiscale features of small infrared targets. Zuo et al [27] designed a multiscale feature fusion pyramid module (AFFPN) with ResNet as the backbone to deal with the loss of targets in the deep network. Wang et al [28] used ResNet as the backbone, used region proposal network (RPN) to extract candidate targets, used FCN [29] network to extract feature map, and used Transformer [30] to determine candidate targets.…”