Abstract: In intensive poultry farming, caged laying hens are highly susceptible to infectious diseases, while conventional manual inspection methods are in low efficiency and delayed responses. This study proposed a lightweight object detection network, PoultryFecesNet-Lite, enabling high-precision monitoring of typical poultry diseases through real-time detection of pathological features from fecal images. It can contribute to early warning of disease infection and precise control. An image dataset was constructed including fecal of hens infected by Coccidiosis, Newcastle disease, and Salmonella, as well as the fecal of healthy hens. Data augmentation was performed using the DiffuseMix diffusion model to alleviate class imbalance and improve recognition performance for underrepresented categories. The proposed deep learning network, PoultryFecesNet-Lite, was rephrased from the YOLO11n framework, incorporating multiple modules to enhance accuracy and efficiency. In which, GSConv lightweight convolution and the VoVGSCSP cross-layer feature fusion module reduced computational redundancy while preserving critical semantic information. MV2Block and MobileViTBlock modules strengthened mid- and high-level semantic feature extraction, improving recognition of overlapping samples and mitigating background interference. Multi-scale feature aggregation was achieved through a DSPPF (Double-SPPF) layer, adapting to fecal targets at varying observation distances. Finally, a layer-adaptive magnitude pruning (LAMP) strategy streamlined the network, reducing parameters and computational cost without compromising detection performance. Meanwhile, the Heatmap Intersection over Union (IoUheat) and Center Offset Distance (D_c) was introduced into the evaluation system to complement conventional object detection metrics. While standard indicators such as mAP and Precision characterized the model’s classification and localization accuracy (i.e., predictive correctness), the interpretability metrics (IoU_heat and D_c) quantified the spatial consistency between the model’s high-response regions and actual lesions (i.e., the rationality of feature focus). Together, these metrics validated the model’s capability to filter out interference and precisely capture pathological features in complex backgrounds, ensuring the robustness and reliability of the PoultryFecesNet-Lite model for early disease warning in intricate farming environments. Extensive tests demonstrated that DiffuseMix augmentation effectively addressed dataset imbalance, enhancing recognition capability for scarce categories. The training results of PoultryFecesNet-Lite presented a mean Average Precision (mAP@0.5) of 92.35%, with 1.47M parameters and 2.20 GFLOPs. Compared with YOLO11n, the parameter amount reduced by 1.14M (43.67%) and the computational cost was reduced by 4.28G (66%), while the mAP@0.5 increased by 0.55%, which achieved a balanced trade-off between accuracy, real-time performance, and resource efficiency. Grad-CAM visualization result confirmed accurate localization of key pathological features in all categories, providing interpretability and reliable focus on disease-specific characteristics. Ablation studies validated the effectiveness of individual modules and their combined contributions. The improved PoultryFecesNet-Lite model simultaneously increased IoU_heat and reduced the center offset distance (D_c) across all categories, indicating that the recognition results not only covered targets more comprehensively but also aligned more precisely with target centers. Specifically, the Ncd class exhibited the most significant improvement in IoU_heat, rising by 0.103, and achieved the greatest reduction in D_c, decreasing by 0.045. The synchronized enhancement in both coverage precision and localization concentration demonstrated the effectiveness of employing the DiffuseMix diffusion model to augment the training samples for the Ncd category. In conclusion, three achievements can be summarized for study, first, a dataset of 9,511 images covering four conditions was built using DiffuseMix to mitigate class imbalance and background interference, boosting model robustness against scarce samples and complex scenes; second, PoultryFecesNet-Lite integrated lightweight modules and pruning to reduce computational complexity while maintaining high accuracy, significantly enhancing deployment feasibility compared to the YOLO11n baseline; third, visualization analysis via heatmap IoU and center offset distance confirmed the model accurately focuses on lesion areas consistent with pathology, thereby improving its interpretability.