In order to improve the accuracy of dead broiler detection in large-scale broiler farms, based on color images and thermal infrared images, two-stage and one-stage dead broiler detection methods for broilers were proposed, respectively. In the two-stage method, the YOLO v11-seg network was firstly used to segment broilers in color images to obtain broiler mask coordinates; then individual broiler thermal infrared images were extracted and classified by using the YOLO v8-cls classification network. In the one-stage method, G-channel replacement fusion images, weighted fusion images, wavelet transform fusion images, and frequency domain transform fusion images were constructed based on color images and registered thermal infrared images. Multi-source fusion image datasets were used to build a dead broiler detection model based on the YOLO v11s object detection network. The results showed that in the two-stage dead broiler detection method, the mAP of broiler instance segmentation was 94.2%, and the classification accuracy of individual broiler thermal infrared images was 99.4%. In the one-stage dead broiler detection method, the model built based on wavelet transform fusion images achieved the highest detection accuracy, with mAP of 93.0%. Compared with the two-stage method, the one-stage detection method had a higher precision rate of 92.3% on the public test set, faster inference speed (6.1 ms/f), and easier to be deployed. Analysis of the temperature distribution of individual broiler thermal infrared images indicated that there were significant differences in body surface temperature distribution between low-age and high-age broilers. The dead broiler detection method proposed can accurately identify dead broilers in the harsh imaging environment under high-density breeding, and it can provide a technical reference for the death detection of other livestock and poultry.