Target Detection in Rural River Image Based on Yolo V5 Model with the Cross-Stitch Unit

Abstract

Due to the complexity of rural river environment, the performance of traditional Yolo V5 model for target detection is insufficient. Considering that the river environment includes river drifts and river slope accumulation, they have different characteristics but have correlation. Therefore, the target detection of river image includes two tasks: identifying the drifts on the river channel and identifying the accumulation on the river slope. In addition, there is no uniform standard to design the shared layer when using deep learning model for multi task learning. The cross-stitch network can decide the best sharing layer through end-to-end learning, and the cross-stitch unit can stitch two networks. Therefore, we introduce the cross-stitch unit into the backbone layer of Yolo V5 for the first time, which makes the YoloV5 network model a multi task learning model. At the same time, we use gamma correction method to preprocess the river drifts image set to improve the detection accuracy of the model. We took field photos of rural rivers and collected similar photos on the Internet to train and evaluate the model. The experimental results show that the average accuracy (map) of the proposed method is 78.6% and 78.0% respectively, which is 2.1 % and 2.3% higher than YoloV5 model, and better than the traditional target recognition model.