Neural network models have been extensively applied in various domains, such as autonomous driving, facial recognition, and machine translation. However, deploying neural networks in scenarios with limited computing power faces challenges such as inefficient training and slow convergence. Therefore, we propose a Lightweight Data Selection Network (LDSN) tailored for constrained computational environments for image classification. Specifically, LDSN employs a scoring model to swiftly identify samples suitable for training the lightweight classification network. It initializes a lightweight scoring network to score and rank samples from the original dataset, then utilizes high-scoring samples to train the classification network, and updates the scoring network according to validation accuracy and model loss. Besides, to address the discrete nature of sample selection, we introduce the policy gradient algorithm for training the scoring model. Finally, we employ a Limited-memory Broyden-Fletcher-Goldfarb-Shanno (LBFGS) optimization algorithm to reduce training epochs of the classification network, thereby enhancing training velocity. Experimental results on standard image datasets demonstrate the effectiveness and efficiency of LDSN.