Abstract: This study aims to estimate the feed intake of chickens in stacked-cage-raising systems. Meat-type breeder chickens were selected as the research object. An improved lightweight version of the YOLOv8n model, named YOLO-FSG, was developed to recognize chicken feeding behaviors. A multivariate linear regression model was established to estimate feed intake using feeding duration and frequency as variables. A dataset of 2043 images was constructed. The data augmentation was applied to the training and validation sets using random rotation, flipping, brightness alteration, and Gaussian blurring. The expanded training and validation set contained 3268 and 408 images, respectively. Meanwhile, the test set contained 204 images. The YOLO-FSG model was achieved in the lightweight performance after three key modifications: 1)In the backbone network, the efficient multi-scale attention (EMA) module, Faster-NetBlock, and C2F (CSPDarknet53 to two-stage feature pyramid network) were combined to create the C2F-FEblock, thus enhancing feature extraction while reducing model complexity. 2) In the neck network, a Slim-Neck design strategy was employed to replace the convolutional layers and C2F modules with GSConv (group shuffle convolution) and VovGSCSP (GSConv spatial cross stage partial) modules, in order to optimize the feature fusion and processing. 3) In the detection head, grouped convolutional modules with two shared parameters were replaced by four convolutional modules to reduce the computational load. The YOLO-FSG model achieved a 97.1% mean average precision at a 0.5 threshold (mAP_0.5) and a 47.4% mean average precision across 0.5-0.95 thresholds (mAP_0.5-0.95), with a model size of 1.94 M, 4.0 G floating-point operations (FLOPs), and a detection time of 3.6 ms. Compared with YOLOv8n, the improved model was enhanced by 0.2 percentage points at mAP_0.5-0.95 and maintained the same mAP_0.5; There was a decrease in the number of parameters, where FLOPs were reduced by 35.3% and 50.6%, respectively; The detection time decreased by 0.2 ms, leading to the more portable model. The improved model outperformed the common object detection models, such as YOLOv5n, YOLOv7n, and YOLOv9t. Feeding duration and frequency were calculated to convert hourly feeding video recordings into sequential image frames. Each frame was analyzed to detect objects using the YOLO-FSG model. The bounding boxes were obtained to detect the entities, in order to complete their coordinates and cage classification. A predefined threshold line was used to delineate the feeding area proximate to the troughs. Once the centroid of a chicken's head fell within this designated feeding area, it was classified as actively feeding. The duration of each feeding episode was determined to count the number of successive frames during feeding activity. An increment was recorded in feeding frequency, whenever there was a pause exceeding 3 s between two feeding episodes. Thus, the feeding frequency and feeding duration were obtained for each cage position within a one-hour period. Correspondingly, the actual feed intake of chickens was measured manually within different cages over one hour. A total of 185 samples were obtained after the removal of unusable data. The training and testing sets were divided at an 8:2 ratio. The feed intake estimation model was achieved in a coefficient of determination (R²) and root mean square error (RMSE) on the test set of 0.9 and 5.91, respectively. In the future, automatic weighing equipment for feed should be utilized to develop more accurate feed intake estimation models. This finding can provide technological support to the fine management of chickens in stacked-cage-raising systems, thereby promoting intelligent development in poultry farming.