The cooling facilities of nuclear power plants are crucial components for transferring the accumulated heat from the reactor to the external environment. However, the complex structure of these facilities makes traditional manual anomaly detection methods inefficient, time-consuming, and poses certain risks to human safety. Therefore, this paper proposes a system for anomaly detection in the cooling source areas of nuclear power plants based on improved multi-modal underwater image recognition. The system utilizes deep learning and image processing techniques to denoise, compensate for illumination, and correct water flow in collected images of the cooling source areas. Wavelet-based noise removal techniques are applied to further enhance image clarity, providing accurate inputs for machine vision algorithms. By training an improved convolutional neural network model, real-time automatic identification and classification of underwater anomalies such as foreign objects and pipeline cracks are achieved, significantly improving the monitoring efficiency and safety of the cooling source areas in nuclear power plants. Comparative analysis with actual cases validates the excellent stability and accuracy of the proposed method in handling underwater operation images in the cooling source areas of nuclear power plants.