Most existing image super-resolution reconstruction algorithms assign equal weights to low-frequency and high-frequency information across convolutional channels. This approach hinders the representation of high-frequency features in images, resulting in blurred edges in reconstructed outputs. To address this issue, we propose a novel image super-resolution reconstruction network called CAGN based on channel attention and gradient information. We introduce a group channel attention mechanism that increases the importance of detailed channels, enhancing the network’s ability to learn high-frequency features. To improve the edge details of the reconstruction image, we extract the image gradient information as prior knowledge in network learning. Meanwhile, we use channel-halving convolution and skip connections to reduce the network’s parameters while deepening the network. Experimental results on five benchmark datasets show that our proposed method achieves higher PSNR metrics than similar methods, ranging from 0.10dB to 3.77dB, while using fewer parameters.