A gradual diffusion attention for the detection of small and complex structures: considering tomato growing point bud as example

Abstract For target detection in complex environments, it is most important to recognize the target object accurately and without error. However, effectively perceiving complex structures and small targets of plants in greenhouse environments remains a challenge in visual recognition tasks. In order to solve these problems, this paper proposes a new attention network, GDA (Gradual Diffusion Attention), based on Deformable DETR. GDA generates feature sampling points by means of gradual diffusion, which enlarges the receptive field while retaining the global context information, effectively enhancing the perception of small targets. In addition, GDA can adaptively focus on the boundaries of complex fine structures by means of dynamic constraints, effectively improving the recognition accuracy in complex environments. Extensive experiments on the greenhouse tomato growing point bud dataset and the COCO benchmark dataset demonstrate the effectiveness of the method. Compared to similar methods, GDA improved the average accuracy of detection of small targets by 10.9% and the overall average accuracy by 7.1%. The code is available at https://github.com/codeimdy/GDANet .