Multistage Algorithm Based on Neural Networks for Preprocessing and Processing of Photo and Video Data from Camera Traps

This paper proposes a technology for analyzing camera trap data using a multistage preprocessing and data processing algorithm based on neural networks. The task of the first stage is separating empty images from nonempty ones. To solve this problem, a comparative analysis of the YOLOv5, YOLOR, and YOLOX architectures was carried out, and the most optimal detector model was identified. The task of the second stage is classifying the objects that were found by the detector. The models EfficientNetV2, SeResNet, ResNeSt, ReXNet, and ResNet were compared. To train the detector model and the classifier, a data preparation approach was developed that consists in removing duplicate images from the sample, and then the method was modified using agglomerative clustering to divide the sample into training, validation, and testing sets. In the problem of object detection, the YOLOv5-L6 algorithm was the best on the collected dataset, with an accuracy of 98.5%; in the task of classifying the found objects, the ResNeSt-101 architecture showed itself to be best of all, with a recognition quality of 98.339% on test data.