The hybrid architecture combining convolutional neural networks with Transformer has attracted much attention for its potential to further enhance the performance of super-resolution reconstruction. However, existing parallel-structured networks typically have independent feature extraction and fusion processes, neglecting the hierarchical interaction and information guidance between heterogeneous modeling paths, making it difficult to simultaneously model low-level spatial details and high-level semantic information. To address this issue, a super-resolution reconstruction network based on progressive heterogeneous feature fusion is proposed. This network alternately performs heterogeneous feature extraction and fusion layer by layer, fully leveraging information at each level. A lightweight adaptive feature fusion module is designed to adaptively select and fuse heterogeneous features using learnable dynamic weights. Additionally, a multi-resolution collaborative context aggregation module is developed, constructing a multi-path branch network with features of large, medium, and small resolutions. Through cross-resolution information interaction, it captures context information. Specifically, the proposed progressive triple-perception residual block enhances feature perception and flexibility through a “local-global-channel” strategy. The cross-resolution reverse projection fusion module builds a learnable difference-projection architecture to achieve dynamic information complementarity and closed-loop interaction among cross-resolution feature maps. Experimental results show that the proposed method achieves the best reconstruction performance on multiple datasets compared with current advanced methods of the same kind.