Abstract Efficiently detecting objects in the complex background at night with low illumination remains a challenge for image processing. To address this issue, this paper proposes LiteCortexNet, a lightweight deep learning object detection model inspired by the visual cortex. The model performs intrinsic image decomposition end-to-end to obtain the illumination-independent reflection component, fuses it with the output result of the depth-wise separable convolutional encoder, and then, sends it to the lightweight detection head for object classification and positioning. Leveraging the channel-wise attention mechanism, our model is optimized for detecting small objects as well as obscured objects. In order to evaluate our method, an image dataset of railway maintenance tools was constructed. Experimental results show that the proposed model achieves 90.56 mAP at 66FPS on this dataset, which outperforms state-of-the-art object detection models such as YoloV4 (Bochkovskiy et al. in arXiv:2004.10934) (82.34 mAP at 45FPS).
展开▼
