Abstract: To improve the recognition and localization performance of infrared object detection algorithms, deep learning technology was applied to infrared image feature analysis, and an infrared object detection model based on feature-layered convolution was proposed. First, the model utilizes traditional image-processing methods to enhance the infrared object information, and a layered convolutional backbone network based on feature keying is designed to extract the infrared object features gradually from shallow to deep. Second, for features in different receptive fields, multiscale feature fusion is achieved by combining a dilated convolution with a channel-space adaptive fusion structure. Finally, based on the fused features, the categories and positions of infrared objects are predicted, and the final objects are selected and fused according to factors such as prediction box confidence, overlap area, and center-point distance. Experimental results on the public FLIR and KAIST datasets show that the proposed method can effectively utilize computational resources to deeply mine key features in infrared images. Compared with similar algorithms, it also demonstrates better recognition and localization performance, making it more suitable for infrared object detection tasks.