Fish behavior recognition can be expected to assess the health status of aquaculture species. Early detection of diseased individuals can also prevent large-scale disease outbreaks. Among them, Takifugu rubripes hold significant economic value in various farmed fish species. However, its abnormal behaviors, such as body tilting, rolling, or vertically floating near the water surface, are often highly similar to normal behaviors under underwater observation. This high behavioral similarity has posed a substantial challenge to current vision systems, frequently leading to misjudgment or false classification. In this study, an abnormal behavior recognition was proposed for Takifugu rubripes using dual correlation feature fusion, termed DCFF-EYSFNet. Target localization and behavior classification were constructed in the framework. Each sub-network was focused on a specific task. The overall training complexity was reduced to avoid the gradient conflicts commonly observed in end-to-end models that simultaneously attempted detection and classification. In target localization, the raw video frames were first preprocessed to extract and fuse the edge contours of individual fish. The posture information of each fish was enhanced to distinguish it from complex underwater backgrounds. Subsequently, an efficient coordination attention module (ECAM) was designed and integrated into the detection network. Two components consisted of: an efficient channel attention block (ECAB) and a coordinated spatial attention block (CSAB). Specifically, ECAB adaptively recalibrated channel-wise feature responses, while CSAB captured long-range spatial dependencies along both horizontal and vertical directions without excessive computational overhead. The most discriminative features of individual fish were focused on to significantly improve the accuracy of fish localization, even in densely populated or partially occluded scenarios. The precise localization provided reliable spatial input for the subsequent image processing. In behavior classification, there were subtle differences between normal and abnormal movements. A dual correlation feature fusion (DCFF) mechanism was introduced to enhance the correlation between temporal and channel dimensions. Spatiotemporal features were effectively fused at multiple hierarchical levels of the network. The DCFF mechanism was composed of the Hjorth and the path feature reconstruction (PFR) module. The Hjorth module was used to capture high-order temporal statistics, specifically the Activity and Mobility parameters, thus representing the stability and instantaneous intensity of fish motion over time. The PFR module was used to separate strong and weak features from the slow and fast pathways of the SlowFast backbone. Redundant background noise was also suppressed to reconstruct a purified feature representation using adaptive weighting. The DCFF, after modules integration, effectively enlarged the feature discrepancy between normal and abnormal behaviors, thereby enabling more accurate and reliable recognition. Furthermore, an underwater dataset was collected from real aquaculture environments, including both clear and turbid water conditions under varying illumination. Both ablation studies and comparative evaluations were performed to validate the effectiveness of each component. The experimental results demonstrate that the DCFF-EYSFNet was substantially improved over the baseline model. Specifically, the accuracy and recall rate increased by 7.8 and 7.6 percentage points, respectively, compared with the baseline SlowFast network. Furthermore, DCFF-EYSFNet achieved superior performance, with F1-score improvements of 4.7 and 5.8 percentage points, respectively, compared with state-of-the-art spatiotemporal action detection models, such as YOWOv2 and ST-GCN. The normalized confusion matrix revealed that the framework correctly identified 97% of normal behaviors and 85% of abnormal behaviors, where only 1% of abnormal samples were misclassified as normal. The rest 14% of missed detections were attributed to severe occlusion or blurred edge features during localization, which was explicitly acknowledged as a current limitation. The DCFF-EYSFNet can be expected to accurately recognize the abnormal behaviors of Takifugu rubripes in complex underwater environments. The findings can provide effective technical support for early disease warning and targeted intervention in precision aquaculture, thereby contributing to minimal economic losses and antibiotic usage more sustainable for fish farming practices. However, the current model can still suffer from relatively high inference latency due to its serial processing pipeline, thus limiting its real-time deployment on resource-constrained edge devices. Network lightweighting, knowledge distillation, and hardware acceleration can improve inference efficiency with high detection accuracy.