A study on the application of background separation preprocessing technique in bird recognition modeling

In order to address the problems of insufficient accuracy of fine-grained bird recognition in complex natural scenes, data scarcity and serious background interference, this project proposes to weaken the environmental interference and emphasize the subject features by the pre-processing method of “Background Removal combined with Noise Replacement”, and then fine-tune the training based on the ResNet-50 Migration Learning; the experiments show that the accuracy of model validation is improved from 75% to 82% after pre-processing, which is significantly better than the unprocessed and conventional data enhancement methods. Experiments show that the model validation accuracy increases from 75% to 82% after preprocessing, and maintains more than 90% recognition effect in the occlusion test and online demonstration, which is significantly better than the unprocessed and conventional data enhancement methods, and provides a low-cost and high-efficiency improvement path for the applications of nature conservation monitoring and electric power bird pest control.