This paper proposes a multi-scale image estimation method based on wavelet transform to effectively remove motion features from multiple videos. An autoencoder with sparsity limit is used to adjust the input signal for effective compression. Effective features are extracted, and the optimal unique vector is learned. An improved convolutional neural network (CNN) is used to recognize weak moving objects. Experiments show that the algorithm achieves high accuracy without large-scale learning samples, with the highest recognition rate of 99.36%, significantly improving upon conventional algorithms. The paper discusses the challenges of identifying weak moving objects in complex images, where they appear as single or few pixels and are susceptible to noise. Deep learning, particularly CNNs, has made significant progress in various fields. However, CNNs face limitations in computing power, network depth, and optimization algorithms when applied to other types of recognition. The paper also reviews existing methods for weak moving object detection, including wavelet signal transformation and time-space non-local similarity, which have limitations in accurately identifying similar targets. The paper introduces an intelligent image recognition system based on the U-net network for efficient feature connection and fusion. The U-net network is used for semantic segmentation, with an input size of 512x512. The network has a U-shape, making it suitable for semantic segmentation in medical images and traffic signs. Convolutional neural networks are also discussed for character recognition, with two-dimensional convolution operations used for image processing tasks. The paper presents a CNN-based multimodal learning algorithm for dim unit target recognition. The algorithm uses fixed features from continuous and non-continuous frames to optimize recognition methods, ensuring high accuracy. Simulation experiments using small-scale ImageNet data show that the algorithm achieves high recognition accuracy, with the highest accuracy at 72 hidden neurons. The algorithm outperforms existing text-based comparison methods in terms of accuracy. The paper concludes that the proposed algorithm can achieve high-precision image recognition without relying on large-scale sample data, improving efficiency and accuracy in weak moving target identification.This paper proposes a multi-scale image estimation method based on wavelet transform to effectively remove motion features from multiple videos. An autoencoder with sparsity limit is used to adjust the input signal for effective compression. Effective features are extracted, and the optimal unique vector is learned. An improved convolutional neural network (CNN) is used to recognize weak moving objects. Experiments show that the algorithm achieves high accuracy without large-scale learning samples, with the highest recognition rate of 99.36%, significantly improving upon conventional algorithms. The paper discusses the challenges of identifying weak moving objects in complex images, where they appear as single or few pixels and are susceptible to noise. Deep learning, particularly CNNs, has made significant progress in various fields. However, CNNs face limitations in computing power, network depth, and optimization algorithms when applied to other types of recognition. The paper also reviews existing methods for weak moving object detection, including wavelet signal transformation and time-space non-local similarity, which have limitations in accurately identifying similar targets. The paper introduces an intelligent image recognition system based on the U-net network for efficient feature connection and fusion. The U-net network is used for semantic segmentation, with an input size of 512x512. The network has a U-shape, making it suitable for semantic segmentation in medical images and traffic signs. Convolutional neural networks are also discussed for character recognition, with two-dimensional convolution operations used for image processing tasks. The paper presents a CNN-based multimodal learning algorithm for dim unit target recognition. The algorithm uses fixed features from continuous and non-continuous frames to optimize recognition methods, ensuring high accuracy. Simulation experiments using small-scale ImageNet data show that the algorithm achieves high recognition accuracy, with the highest accuracy at 72 hidden neurons. The algorithm outperforms existing text-based comparison methods in terms of accuracy. The paper concludes that the proposed algorithm can achieve high-precision image recognition without relying on large-scale sample data, improving efficiency and accuracy in weak moving target identification.