PoultryFecesNet-Lite: A disease monitoring model for fecal image identification in intensive poultry farming of laying hens

Caged laying hens are highly susceptible to infectious diseases in intensive poultry farming. While conventional manual inspection cannot fully meet the needs of large-scale farming, due to the low efficiency and delayed responses. In this study, a lightweight deep learning network, PoultryFecesNet-Lite, was proposed for high-precision monitoring of typical poultry diseases, based on the identification of pathological features from fecal images. An image dataset was constructed, including the fecal of hens infected with Coccidiosis, Newcastle disease, and Salmonella, as well as the feces of healthy hens. Data augmentation was performed to alleviate class imbalance for recognition performance on underrepresented categories using the DiffuseMix diffusion model. PoultryFecesNet-Lite was rephrased from the YOLO11n framework. Multiple modules were incorporated to enhance accuracy and efficiency. Among them, GSConv lightweight convolution and the VoVGSCSP cross-layer feature fusion module reduced computational redundancy to preserve critical semantic information. MV2Block and MobileViTBlock modules strengthened the extraction of mid- and high-level semantic features. Overlapping samples were recognized to avoid background interference. Multi-scale feature aggregation was achieved in a DSPPF (Double-SPPF) layer, suitable for fecal targets at varying observation distances. Finally, a layer-adaptive magnitude pruning (LAMP) was applied to streamline the network, thus reducing parameters and computational cost without compromising detection performance. In addition, the Heatmap Intersection over Union (IoU_heat) and center offset distance (D_c) were introduced into the evaluation system to enhance conventional object detection metrics. Standard indicators, such as the mean Average Precision and Precision, were employed to represent the classification and localization accuracy (i.e., predictive correctness), while the interpretability metrics (IoU_heat and D_c) were quantified for the spatial consistency between the model’s high-response regions and actual lesions (i.e., the rationality of feature focus). These metrics were validated to filter out the interference and then precisely capture pathological features in complex backgrounds. The training results of PoultryFecesNet-Lite presented the mAP@0.5 of 92.35%, with 1.47 M parameters and 2.20 G FLOPs. The parameter amount and computational cost were reduced by 1.14 M and 4.28 G, respectively, compared with YOLO11n, while the mAP@0.5 increased by 0.55 percentage points. Grad-CAM visualization result confirmed that there was the 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 IoU_heat increased simultaneously, whereas the center offset distance (D_c) decreased for all categories. The recognition bounding box also covered more targets to precisely align with the target feature centers. In conclusion, three achievements were summarized: (1) A dataset was constructed to reduce class imbalance and background interference using DiffuseMix, thereby consisting of 9511 images under four conditions; (2) PoultryFecesNet-Lite model was integrated with multiple modules and pruning to reduce computational complexity for the high accuracy; (3) Visualization analysis via heatmap intersection over union and center offset distance confirmed that the model accurately focused on pathological lesion areas, indicating its interpretability. The findings can also contribute to early warning of disease infection and precise control in modern poultry farming.