计算机科学
红外线的
分辨率(逻辑)
人工智能
光学
物理
作者
Chenfan Sun,Guangming Dai,Maocai Wang,Lei Peng,Xiaoyu Chen,Zhiming Song
标识
DOI:10.1016/j.engappai.2024.107924
摘要
Infrared weak and small target detection algorithms have important applications in the field of infrared remote sensing. Since the small pixel proportion in the imaging plane and a lack of distinctive features, achieving accurate and rapid detection for infrared weak and small targets remains a highly challenging problem. Building upon this, we proposed a high-resolution infrared weak and small target detection(HR-IWSTDet) model. 1.HR-IWSTDet constructs a backbone with multi-resolution subnetwork in parallel, using the output feature map from the 2× high-resolution subnetwork as head to retain fine-grained features, ensures that more positive samples are attended to during the label assignment process. 2. Introduced the channel-splitting attention(CSA) Block, which utilizes the Cross-Resolution Spatial Attention Module (CRSAM) and Single-Resolution Channel Attention Module (SRCAM) to replace two 3 × 3 convolutions in the BasicBlock, enables information flow in both spatial and channel domains, significantly reducing model parameters and inference time. 3. Adopted an enhanced coordinate representation by decoupling the horizontal coordinate −x and vertical coordinate −y into two separate one-dimensional vectors. These vectors are adjusted by a scaling factor s to encode the center point coordinates with a finer measuring unit. One-dimensional vector calculations use Gaussian kernel functions, considering spatial correlations between adjacent labels. Moreover, we built the IWSTD dataset, which consists entirely of infrared weak small target samples. Experimental results on this dataset show that HR-IWSTDet has a parameter count of only 2.60M, achieving an Average Precision (AP) of 84.85%. The inference time for a single frame is as low as 0.015 s. HR-IWSTDet also outperforms existing methods on the public IASTD dataset. Experimental data validates the effectiveness and generalization of the proposed approach.
科研通智能强力驱动
Strongly Powered by AbleSci AI