Abstract: In recent years, the construction and operation of large-scale radio telescopes such as the Square Kilometre Array (SKA), the Five-hundred-meter Aperture Spherical Telescope (FAST), and the Next Generation Very Large Array (ngVLA) have ushered radio astronomy into the era of big data. A radio source detection method based on Convolutional Neural Networks (CNNs) is proposed to address the demands of processing massive amounts of data. Targeting the characteristics of weak signal recognition and large range of scale structure changes in radio images, a deep learning framework combining a dual-channel denoising model with a feature pyramid network is constructed. First, dilated convolutions and residual learning are employed to enhance the signal-to-noise ratio (SNR) of the images. Then, multi-scale feature fusion is utilized to achieve precise localization of both point sources and extended sources. The experiments were conducted using the dataset from the first edition of the Scientific Data Challenge (SDC1). The results demonstrate that the proposed method achieved detection accuracies of 97.7%, 94.6%, and 93.8% for images with integration times of 1000 hours, 100 hours, and 8 hours, respectively. Notably, the method's ability to detect weak sources in the 8-hour images showed a significant improvement of 53.2% compared to the best-performing team. Furthermore, the search results above the noise level achieved both reliability and completeness exceeding 90%, effectively addressing the challenge of detecting weak signals in high-noise environments.