A study of rainy ibis detection based on Yolov7-tiny

Abstract

The YOLOv7-tiny algorithm does not achieve high detection accuracy for crested ibis in rainy environments. Therefore, we developed a rainy day crested ibis target detection algorithm based on YOLOv7-tiny. Firstly, the RainMix method is used to simulate the rainy day shooting data to synthesise a set of ibis dataset which is closer to the real environment. Then, the k-means algorithm is applied to re-cluster the predicted anchor frames to improve the approximation between the predicted and real frames in the output. Finally, an efficient hybrid attention mechanism (E-SEWSA) is developed and integrated into a lightweight efficient layer aggregation network, while a dense residual network reconstruction module is utilised to improve the detection accuracy of the model. In the PAN+FPN structure, the context information fusion capability of the feature aggregation part of the network is enhanced by integrating the CARAFE module instead of the up-sampling module, so as to improve the model detection accuracy. After experimental verification, the algorithm proposed in this paper has better results in rainy day ibis detection.