<p>Convolutional neural networks (CNNs) have made significant advances in various vision tasks but are limited by their substantial computational requirements. Many researches focus on eliminating redundant information from the feature maps to achieve lightweight CNNs. However, this approach may unintentionally discard still useful detailed information. To address the issue, we introduce maximum dissimilarity channel complementary reconstruction (MDCR) module, which is specifically designed to repurpose perceived redundant channel information as supplementary data. The MDCR module includes three units: an information-aware separation unit that bifurcates feature maps into information-rich branch and information-redundant branch, applying depth-wise convolution to redundant branch; a complementary reconstruction unit employs the rational utilization of redundant information, whereby channel matching is conducted based on the principle of maximum dissimilarity; and a circular convolution channel attention unit that enhances feature representation. The efficacy of the MDCR module has been demonstrated through rigorous experimentation on datasets such as MS COCO, CIFAR, and Komatsuna. Compared to standard CNNs, MDCR shows a substantial reduction in computational demands by approximately 87% while achieving better accuracy across all datasets.</p>