Accurate identification of solar radio bursts (SRBs) is of great significance for solar physics research and space-weather forecasting. Most existing studies focus on the mere detection of SRB occurrence or the identification of a single class (e.g., Type III bursts), which fails to meet the demand for precise detections of various solar radio bursts. Additionally, current mainstream SRBs detection models often employ complex architectures and redundant parameters, resulting in low computational efficiency. To address these limitations, we constructed a spectrogram dataset based on the e-CALLISTO platform, comprising Type II, Type III, Type IV, and Type V bursts. The dataset contains 8752 images with 10,822 annotated instances, where samples of types IV and V are incredibly scarce. To overcome the challenge of pretraining with few-shot classes, this paper proposes a pretraining method that integrates a stable diffusion generative model with a self-supervised learning strategy, effectively enhancing the model's learning capability for few-shot classes. Building on this, this paper presents a detection model for various solar radio bursts, VitDet-SRBs (Vision Transformer Detector for Solar Radio Bursts), which incorporates a channel attention mechanism into the feature fusion module to enhance performance while controlling model complexity. Experimental results show that VitDet-SRBs achieve an average precision at a single Intersection-over-Union threshold of 0.50 (AP@50, AP with IoU = 0.50) of 81.2% on the SRBs dataset, outperforming existing mainstream methods in both precision and recall. This study not only provides a novel approach for efficient detections of various solar radio bursts but also offers a feasible solution for other few-shot astronomical data processing problems, with broad application prospects.