Infrared Small Target Detection Based on Local Contrast-Weighted Multidirectional Derivative

Robust Unsupervised Multifeature Representation for Infrared Small Target Detection

“…However, the customized feature extraction model is difficult to adapt to different detection scenarios, while the limited feature representation capability limits further improvement of detection performance. Some other researchers have attempted to introduce more features to enhance the feature representation capability, such as scale [52], gradient [44] and weighted information [60]. However, limited by the accuracy of feature extraction, they are prone to cause more missed detections while reducing false detections.…”

Real-Time Infrared Small Target Detection With Nonlocal Spatial-Temporal Feature Fusion

“…Due to the limitations of classic methods, many researchers began to solve the problem based on convolution neural networks (CNN) or transformers. They [1] [2] [13] [14] [26] [27] performed much better than classic methods even on a dataset that is not overlapped with the training subset [2]. Though the deep learning based methods showed powerful abilities in object detection, it was a certain waste of computing resources for smaller target detection by simply adopting general frameworks.…”

“…For lightweight model building, besides efficient structures like depthwise separable or group convolution in some mobile networks [30] [31], it was common to clip the unnecessary branches [32] [33] [12] or reuse some backbones or modules [34] to make the models efficient. Especially for smaller objects or ISTs, many researchers also found the importance of shallow features [2] [35] [36]. For example, Sun et al [35] designed a simple Multi-Receptive Field Extraction (MRFE) module to further extract features with different receptive convolutions.…”

“…They thought the shallow layers were suitable enough and the MRFE was light with only 12 layers. Different from MRFE, some other researchers [2] [36] extracted features from low to deep layers with a few convolutions. Widely adopted residual direct skip and dimension concatenation made the convolution combinations abundant, leading to features with adequate receptive fields and properties while the computations were controlled to small amounts.…”

“…1. Since segmentationbased methods [1] [14] are too time-consuming and cannot be applied on CPU-only or some source-limited machines for real-time processing, we build the NLMF-Net on the object level in a regression manner [2] [9] [12]. Besides, ISTs are sparse on the spatial dimension.…”