Fine-grained object recognition represents a practical requirement for intelligent interpretation of high-resolution remote sensing imagery. Existing research primarily concentrates on the detection of stumpy targets. Nevertheless, slender objects characterized by a height significantly exceeding their length or width are also common in practical applications. Current research is inadequate to tackle the challenges presented by slender targets, and there is an urgent need for effective methodologies to address this issue. To this end, this paper proposes a model, named Generalized Adaptive Rotation Faster R-CNN (GA-RFRCNN). The GA-RFRCNN optimizes feature representation across multiple scales by integrating selective enhancement feature pyramid network (SE-FPN). Besides, it introduces an enhanced rotation region proposal network (ERRPN) to enhance the object localization. Furthermore, a dynamically adjusted training process is used to handle difficult-to-detect samples by introducing the adaptive slide loss (ASLoss). We conduct extensive experiments on the transmission tower custom dataset (TT-OBB) and the HRSC2016 dataset, and the results show that our model achieves significant improvements in recognition accuracy and oriented bounding box detection.