TFFM-YOLO: An Efficient High-Altitude Multi-Target Detection Algorithm on UAV Infrared Photography

Infrared detection and tracking have emerged as pivotal aspects in the domain of object detection in recent years. However, detecting small objects continues to pose a significant challenge due to their diminutive size and the tendency for indistinct feature details to result in their omission. Moreover, information pertaining to small objects is often lost during downsampling operations. Deep learning-based detection methods have been employed to address these challenges. In this work, we propose an efficient aerial target detection algorithm, TFFM-YOLO, designed to increase detection accuracy and enhance image details within convolutional networks. Firstly, we integrate an Enhanced Receptive Field Augmentation Module (ERFB) into the backbone, aiming to mitigate background interference and amplify target features. Furthermore, we have developed a Triple Feature Fusion Module (TFFM) while leveraging three feature maps from different levels for additional fusion before their input into the detection network. Finally, the normalized Wasserstein distance is employed to optimize the balance of small target thresholds, thereby enhancing the probability of accurate bounding box detection. On the Hit-UAV dataset, our method achieves an 11.5% improvement in [email protected] with minimal changes in the number of parameters compared to the original algorithm.