To address the issue of texture detail blurring in single-scale sampled compressed sensing image reconstruction at low sampling rates and achieve a lightweight reconstruction network， this paper proposes a filter pruning based multi-scale compressed sensing image reconstruction network. In the sampling phase， the image is linearly decomposed by convolution， and then fused with the input image and different scale decomposition features to obtain the compressed sensing measurements. In the reconstruction phase， a coordinate attention based multi-scale dilated residual module is designed， which incorporates positional information into channel attention to enhance the feature learning ability of the network. Moreover， by calculating the entropy of the feature map to judge the importance of the filters， the less important filters is pruned to achieve the purpose of compressing the model. Training and testing are carried out on datasets such as DIV2K， Set5， BSDS68 and Urban100. The experimental results show that the algorithm proposed improves the Peak Signal-to-Noise Ratio （PSNR）， Structural Similarity（SSIM）， and image visual effects. For instance， with a sampling rate of 4% and a test set of Set14， the proposed algorithm improves the PSNR of the reconstructed image by 4.17 dB and 2.39 dB， respectively， compared with CSNet+ and FSOINet， resulting in clearer texture details. Under the premise of slightly reducing the reconstruction effect， a lighter model was obtained， which improved the reconstruction speed.