LRFE-YOLO: Low-Rank Feature Enhancement for YOLO in Few-Shot Industrial Meter Fine Needle Detection

Traditional mechanical needle industrial meters are widely used in the industrial field due to their good stability. As a representation of typical case of small targets, the detection of fine needles in the dial is an important research field in deep learning object detection. However, existing methods may lose a large amount of contextual information in feature extraction and spatial transformation of input data layer by layer. In addition, deep learning requires a large amount of training data, but the collected data of industrial meters is very limited and similar. Therefore, this article proposes a fine needle detection method based on LRFE-YOLO, which solves the problems of small targets and sample scarcity by embedding attention mechanisms into the object detection model. Firstly, LRFE (Low Rank Feature Enhancement) utilizes a rectangular self attention mechanism in the horizontal direction to capture nonlocal similarity in the spatial domain. Secondly, the new channel attention mechanism utilizes the information redundancy of three-dimensional deep features to extract global low rank characteristics that better represent small target information. Thirdly, LRFE-YOLO chooses to add transformer structures containing LRFE to the end of the backbone section of the You Only Look Once (YOLO) framework, making it easier to capture global features. The experimental results show that LRFE-YOLOv8s outperforms YOLOv8s in mAP50-95 on small sample single target (needle) annotated datasets by 3.4%. LRFE-YOLOv8s has mAP50-95 and needle single AP values 2.9% and 3.1% higher than YOLOv8s on multi-objective (including needle) annotated datasets, respectively. Moreover, the LRFE module outperforms common attention mechanisms in improving YOLOv8. The code can be searched in https://github.com/AspiringOliver/LRFE-YOLO.