To improve feature selection in convolutional neural networks for image classification, and to solve the problems of limited feature expression caused by fixed weight calculation in traditional convolution and difficulty in focusing on critical regions, an image classification network of attenuated convolution (ACNet) is proposed. First, an attenuated convolution (AConv) module is designed, which dynamically adjusts convolution kernel weights through a neuron signal attenuation mechanism, constraining overly strong feature responses while preserving necessary weak features, thus achieving biologically inspired feature selection. Second, the AConv module is applied to shallow layers and residual blocks, enabling the network to adjust feature expression intensity at different depths according to feature importance, thereby enhancing feature expression capability. Finally, a dynamic attenuation feature focusing (DAFF) module is designed and embedded at the end of the residual path, which calibrates key region features in the compressed feature space and globally optimizes the fused composite features, effectively improving the accuracy of feature selection and the network's classification capability. Experimental results show that, compared with existing mainstream models, the proposed network achieves a better balance between feature selection and feature integrity, with stronger robustness and classification performance.