Infrared target detection algorithm based on multi-scale edge information extraction

Abstract In infrared target detection, infrared images suffer from problems such as low resolution, low signal-to-noise ratio, and poor contrast. These issues make it difficult to clearly present the edges and details of targets, resulting in insufficient extraction of multi-scale edge information and fine-grained features. Meanwhile, infrared targets exhibit diverse scales, complex features, and are prone to motion blur, which further increases the difficulty of feature extraction and leads to low detection accuracy and serious missed detections. To address these challenges, we propose a Context Edge MultiScale Fusion- You Only Look Once (CMF-YOLO) algorithm for infrared target detection. Firstly, to address the issues of insufficient extraction of multi-scale edge information and excessive background interference, a cross stage partial-multi-scale edge information selection module is designed within the backbone network and neck of the model. This module enables the model to select the features most relevant to the target from multi-scale edge information, effectively reducing background interference and thus contributing to more precise target localization. Secondly, the conventional spatial pyramid pooling-fast (SPPF) structure employs a static pooling scale during the process, which hinders its capacity to adapt to the varied target scales and intricate characteristics inherent in infrared images. To address this issue, we propose a module called feature pyramid shared dilated convolution as an alternative to SPPF, which combines shared dilation convolution with the feature pyramid structure to effectively improve the accuracy and efficiency of target detection. In addition, a small target detection layer P2 is introduced to improve the detection accuracy of small targets. Moreover, Inner-SIoU is adopted as a new localization regression loss function to enhance the learning capability for small target samples and accelerate the convergence of regression bounding boxes. Comparative experiments are conducted on the Alpha Track dataset captured by iRay Technology, as well as on the publicly available FLIR and NEU-DET datasets. The results demonstrate that the CMF-YOLO algorithm achieves an mAP@0.5 of 92.3% on the Alpha Track dataset, 86.6% on the FLIR dataset, and 87.9% on the NEU-DET dataset. These findings confirm the applicability and high accuracy of the proposed method in infrared target detection tasks.