KTnet: Hazy weather object detection based on knowledge transfer

Abstract

The current method to address the reduced accuracy of target detection algorithms in hazy weather scenes is mainly to first use image dehazing algorithms to restore hazy images, and then input the restored images into target detection algorithms to obtain detection results. However, the images restored by the image dehazing model deviate from real clear images, and do not completely recover the features required by the target detection algorithm, thus limiting the improvement of the detection accuracy of the target detection model. This paper proposes a hazy weather target detection algorithm based on large convolution kernels and knowledge transfer (KTnet). First, a large convolution attention dehazing module is embedded into the backbone network of faster R-CNN to form a dehazing backbone network. Considering the high-dimensional features of the deep backbone network, a lightweight fusion attention module is designed. A loss function is also designed and the adapter model is employed to devise training methods for knowledge transfer and fine-tuning. Extensive experimental results on various hazy weather target detection datasets show that KTnet has achieved significant effectiveness.