Abstract: Detecting residual bait particles in complex aquaculture environments presents significant challenges due to their small size, low contrast against the background, and various interferences such as illumination variations and background noise. These limitations often result in suboptimal performance when using conventional detection methods. To address these issues, this paper proposes YOLO-Bait, an enhanced detection model built upon YOLOv8n that incorporates optimized local and global feature extraction mechanisms. First, a central difference convolution (CDC)-based module, termed C2f-CDC (CSP-Bottleneck with two CDC layers), was developed to replace the original C2f module in the baseline architecture. This modification enhances feature representation by aggregating both intensity and gradient information from the feature maps, thereby improving the model's sensitivity to residual bait particles. Second, a multi-scale feature aggregation (MSFA) module was constructed by integrating atrous convolution with a spatial pyramid structure, replacing the original spatial pyramid pooling fast (SPPF) module in YOLOv8n. This design effectively mitigates semantic gaps across different feature levels, facilitates cross-level feature interaction, and preserves critical low-level spatial details. Third, to improve feature fusion capability in the neck while simultaneously reducing parameter count and increasing inference speed, the neck architecture was redesigned using an aggregate-and-distribute mechanism. This redesign enables global integration of multi-level features and injects the fused information back into each layer. Additionally, replacing large detection heads with smaller ones improved detection accuracy for fine bait particles. Finally, the scaled intersection over union (SIoU) loss function was adopted, incorporating both distance and shape constraints to enhance bounding box robustness. Comparative experiments conducted on residual bait datasets collected from recirculating aquaculture of nibea albiflora and pond culture of litopenaeus vannamei demonstrated that YOLO-Bait achieved a precision of 94.6%, recall of 93.7%, mean average precision at IoU (intersection over union) 0.5 of 95.7%, and mean average precision across IoU thresholds 0.5-0.95 of 65.4%. These results represent improvements of 2.6, 8.6, 4.1, and 7.6 percentage points, respectively, over the baseline YOLOv8n model. The model size and parameter count were reduced to 1.14 MB and 2.7 M, corresponding to decreases of 62.1% and 57.8%, respectively, compared to YOLOv8n. Compared with mainstream models including YOLOv9n, Gold-YOLO, YOLOv11, and YOLOv12, as well as existing bait detection models such as YOLOv5s-CAGSDL, YOLO-feed, and YOLO-BaitScan, YOLO-Bait achieved the mean average precision at IoU 0.5 improvements of 3.1, 0.8, 1.1, 0.9, 1.6, 2.3, and 1.9 percentage points, respectively. Moreover, the C2f module optimized with the CDC strategy demonstrated superior performance in detecting small bait particles compared to alternative convolutional strategies, including Partial Convolution (PC), Deformable Convolution (DCN), Strip Convolution (SC), and Pixel Difference Convolution (PDC), with mean average precision at IoU 0.5 improvements of 1.1, 0.9, 1.1, and 2.0 percentage points, respectively. Visualization of heatmaps generated by different convolutional strategies revealed that the CDC-optimized C2f module produced feature maps that effectively focused on bait particle regions while successfully suppressing background interference. Ablation studies confirmed the individual contributions of each proposed component, demonstrating that the C2f-CDC module, MSFA module, redesigned neck architecture, and SIoU loss function collectively improved both detection accuracy and inference efficiency. Additional experiments on mobile devices showed that the deployed YOLO-Bait model achieved a detection rate of 92.1% for residual baits, which is 4.8 percentage points higher than that of the baseline YOLOv8n model, while maintaining a comparable detection frame rate. These findings indicate that YOLO-Bait enables rapid and accurate localization of underwater residual bait particles, providing reliable technical support for quantitative bait assessment in aquaculture operations and contributing to the advancement of smart fisheries management systems.