To address wide catch-size variation, image degradation, severe occlusion, and long-tail class distribution in onboard electronic monitoring images, this paper proposes Fishery-YOLO26, an end-to-end catch detection framework. Based on YOLO26s-P2, it uses the Marine-DFA module to introduce shared image context into multiscale channels and spatial calibration while controlling feature perturbations through bounded residual modulation. The Fish-STAL module adjusts candidate ranking and positive-sample quotas according to class frequency, object scale, and overlap. Experiments show that Marine-DFA mainly provides conditional calibration in degraded scenes, while label-assignment statistics further confirm the practical effect of Fish-STAL on candidate ranking and supervision density for difficult targets. Results on the Fishnet dataset demonstrate improved performance on small targets, tail classes, and degraded scenes while maintaining real-time inference.