Multi-Layer Feature Extraction Object Detection Based on Deep Forest

With the emergence of various large-scale deep learning models, in remote sensing images, the object detection effect is also plagued by complex calculations, high costs, and high requirements for data sets and parameters. Based on simple shallow structures, it is impossible to complete more abstract feature expressions, especially the extraction and fusion of multi-level information. Considering together three key factors affecting deep learning models: layer-by-layer processing, deep feature information extraction, and dependence on datasets. This paper proposes a multi-layer feature extraction object detection method based on deep forest (MLFEDF). The backbone network uses a basic structure to extract multi-scale features, and we design a spatial position information processing module to reduce information loss. Then, the feature refinement process from shallow to deep layers is designed to achieve multi-scale refined object representation. Finally, drawing on the idea of deep forest, multi-level features are jointly utilized to complete the final classification regression. Our model is compared with the existing seven object detection models on the DOTA dataset. The experimental results show that the mAP detection index of MLFEDF reaches 79.15%, which is better than other networks. At the same time, the discussion experiments on parameters and small datasets show that our model can reduce the model's requirements on datasets and parameters, and the accuracy loss is only 3.69%.