Infrared images of caged chickens can provide valuable insights into their health status. Accurately detecting and segmenting individual chickens in these images is essential for effective health monitoring in large-scale chicken farming. However, the presence of obstacles such as cages, feeders, and drinkers can obscure the chickens, while clustering and overlapping among them may further hinder segmentation accuracy. This study proposes a Mask R-CNN-based instance segmentation algorithm specifically designed for caged chickens in infrared images. The backbone network is enhanced by incorporating the CBAM within this algorithm, which is further combined with the AC-FPN architecture to improve the model’s ability to extract features. Experimental results demonstrate that the model achieves average AP and AR¹⁰ values of 78.66% and 85.80%, respectively, in object detection, as per the COCO performance metrics. In segmentation tasks, the model attains average AP and AR¹⁰ values of 73.94% and 80.42%, respectively, reflecting improvements of 32.91% and 17.78% over the original model. Notably, among all categories of chicken flocks, the ‘Chicken-many’ category achieved an impressive average segmentation accuracy of 98.51%, and the other categories also surpassed 93%. The proposed instance segmentation method for caged chickens in infrared images effectively facilitates the recognition and segmentation of chickens within the challenging imaging conditions typical of high-density caged environments, thereby contributing to enhanced production efficiency and the advancement of intelligent breeding management.

Given the issues in natural orchard environments, such as large-scale variations of defective citrus fruits, weak texture boundaries, strong illumination changes, branch and leaf occlusion, and significant background interference, this paper constructs a lightweight detection model, RT-DETR-DCEA, based on RT-DETR-R18. This model is improved through four aspects: “fine-grained defective feature extraction—multi-scale feature fusion—up-sampling detail recovery—global feature interaction for noise suppression”. First, a Dynamic Hybrid Convolution Module (DIMB) is introduced into the backbone network, drawing on the ideas of Inception-style multi-branch depthwise convolution and MetaFormer residual mixing. It extracts local textures of various forms through square convolution, horizontal strip convolution, and vertical strip convolution, and utilizes dynamic branch weights to enhance the model’s adaptability to irregular defects such as lesions, mildew, and external damage. Second, a Content-Guided Attention Feature Fusion Network (CGAFN) is designed in the neck network, which achieves adaptive fusion of low-level detail features and high-level semantic features through channel attention, spatial attention, and pixel-level fusion weights. Next, a lightweight upsampling enhancement module called EUCB-SC is constructed, which introduces channel rearrangement and Shift spatial offset into the efficient upsampling convolutional structure to enhance the local spatial interaction capability of upsampled features with low parameter overhead. Finally, adaptive sparse self-attention is introduced into the AIFI module to form AIFI-ASSA, which suppresses irrelevant background interactions through a sparse attention branch and retains necessary contextual information through a dense attention branch. The experimental results demonstrate that on a dataset containing four categories of citrus images—healthy, diseased, moldy, and severely externally damaged—RT-DETR-DCEA achieves 92.1% Precision, 86.1% Recall, and 91.8% mAP@50, with a parameter count of 1.477 × 10⁷ and an inference speed of 81 FPS. Compared with the original RT-DETR-R18 and various YOLO series models, this method strikes a favorable balance among detection accuracy, recall capability, and model lightweightness. This paper also discusses limitations such as data scale, ratio of private data, single training result, and insufficient validation on edge devices, providing a basis for subsequent cross-regional data validation and real-world deployment testing.

Citrus fruit size is a core quality attribute that determines commercial value and market acceptance. In postharvest processing, rapid, accurate, and non-destructive diameter estimation is an essential prerequisite for automated grading, yet existing computer vision methods are difficult to deploy on factory edge devices. To address this problem, this paper proposes PMDS-YOLO, a lightweight deep learning model for non-destructive citrus fruit diameter estimation and automated quality grading. The model incorporates four key improvements: width factor compression to 0.125 to eliminate channel redundancy; adoption of a MobileOne backbone embedded with the proposed multi-scale channel–spatial residual attention module (MCSR), which reduces parameter count while enhancing multi-scale feature extraction and background discrimination; introduction of dynamic snake convolution (DySnakeConv) to improve the feature extraction module for better detection of irregular fruit contours; and adoption of soft non-maximum suppression (Soft-NMS) to reduce missed detections of densely overlapping targets. Experimental results on a self-built citrus dataset show that PMDS-YOLO achieves only 0.58 M parameters and 86.41% mAP@0.5:0.95, representing a 76.4% reduction in parameter count and a 0.99 percentage point improvement in mAP over the baseline YOLOv13n. In the diameter estimation stage, a quadratic polynomial regression calibration model reduces the MAE to 2.15 mm (R² = 0.85, RMSE = 3.16 mm) and achieves an overall grading accuracy of 57.0% in the grading evaluation, indicating that while the method demonstrates potential in diameter estimation accuracy, further optimization is required for grading performance. The proposed system provides a low-cost, edge-deployable solution for preliminary fruit size assessment and small-scale sorting applications.