Automatic bird detection represents one of the most critical technical challenges in ornithological monitoring systems, which are relevant for scientific wildlife observation, biodiversity assessment, and practical applications in agriculture and environmental management. Modern monitoring systems require high accuracy under real-world imaging conditions; however, automatic detection of birds is complicated by the presence of small and low-contrast objects embedded in complex and highly detailed natural scenes. An additional challenge is the high intra-class variability, which arises from the diversity of bird species, varying viewpoints, and differences in object size, both due to species-specific morphology and varying distances to the camera.This study is dedicated to the development of an effective method for detecting small and low-contrast objects in individual frames of a video stream. The proposed solution is based on a modified SSD-ADSAR architecture enhanced with a dual-stream attention mechanism. On the test dataset, the model achieved mAP@0.5 = 0.876 and mAP@0.5:0.95 = 0.645. The use of synthetically augmented data helped to mitigate the background-type imbalance and improved the model's robustness under complex visual conditions. The practical significance of this work lies in its applicability to real-time ornithological video monitoring systems, as well as to nature conservation, agricultural automation, and scientific ornithological research. The developed method is tailored to typical conditions of ornithological monitoring (such as small, fast-moving objects and cluttered natural backgrounds), and it outperforms existing solutions designed primarily for detecting artificial airborne objects in terms of detection accuracy. © 2025 FRUCT Oy.