10k Is Enough: An Ultralightweight-Binarized Network for Infrared Small-Target Detection

The widespread deployment of infrared small-target detection (IRSTD) algorithms on edge devices necessitates the exploration of model compression techniques. Binarized neural networks (BNNs) are distinguished by their exceptional efficiency in model compression. However, the small size of infrared targets introduces stringent precision requirements for the IRSTD task, while the inherent precision loss during binarization presents a significant challenge. To address this, we propose a binarized IRSTD network (BiisNet), which preserves the core operations of binarized convolutions while integrating full-precision features into the network’s information flow. Specifically, we propose a dot binary convolution (DB Conv) to retain fine-grained semantic information in feature maps while still leveraging the binarized convolution operations. In addition, we introduce a smooth and adaptive dynamic softsign (DySoftSign) function, which provides more comprehensive and progressively finer gradients during backpropagation, enhancing model stability and promoting an optimal weight distribution. The experimental results demonstrate that BiisNet achieves a mean intersection over union (mIoU) of 68.23% with only 10k parameters, delivering an 11.56% improvement over the previous best binary neural network while remaining competitive with state-of-the-art (SOTA) full-precision models.