In light field(LF) imaging technology, both the direction and intensity of light rays are recorded by a special four-dimensional(4D) imaging structure, thereby enabling depth information to be obtained and the relative positions of captured objects to be acquired. Consequently, unique advantages are demonstrated across multiple three-dimensional(3D) application domains. Microlens array-based LF cameras are confronted with the challenge of low spatial resolution because the device - a single detector - is limited in size and simultaneous spatial and angular sampling is required. Currently, numerous non-blind super-resolution(SR) algorithms have been proposed to enhance spatial resolution, where the blur kernel is provided as a known prior input to the network. However, these blur kernels are configured with a single type, and thus remain limited when applied to real-world LF images suffering from complex degradations. Therefore, to address the limitation of non-blind algorithms in handling complex degradation blur, a full-focused blind light field super-resolution(LFSR) network based on deep kernel estimation, named the Depth-Constrained Kernel Estimation Network (DCKE-Net), is constructed in this paper. An iterative optimization architecture is adopted. A depth-constrained spatially variant (SV) kernel estimation module is designed. A feature interaction module is introduced to extract contextual blur information. A parallel window transformer structure is proposed to process spatial and angular features separately for attention calculation. Finally, the network's capability in detail restoration and multi-defocus blur removal is validated through simulation and ablation experiments, providing a blind solution for the full-focus SR task.