In response to the problem of low depth image resolution and the inability of existing super-resolution convolutional networks to be deployed on resource-constrained heterogeneous multi-core processors, this paper proposes a lightweight super-resolution convolutional network named Tiny-FSRCNN. The network incorporates depthwise separable convolution modules to replace certain traditional convolutional layers, and employs low-rank decomposition techniques to decompose large convolutional kernels within the deconvolution layers, thus achieving a lightweight design for the super-resolution convolutional network. Furthermore, this study introduces a hybrid loss function combining Mean Squared Error (MSE) and Structural Similarity Index Measure (SSIM), enhancing the model’s ability to reconstruct image details. Comparative experiments show that while maintaining excellent image super-resolution performance, Tiny-FSRCNN reduces its parameter count from 12,464 to 4,485 and the model size from 0.04886MB down to 0.01863MB. On the same CPU, the inference speed was reduced from 0.7514 s to 0.5599 s.