APCA-Net: Adaptive object detection in rainy weather based on principal component analysis

Since rain has a negative impact on the image processing, traditional image-based target detection algorithms perform poorly on rainy days. The traditional sparse coding based rain-removal method requires the selection and adjustment of many parameters, which requires professional manual adjustment and optimization. Additionally, these methods demand significant computational load and resource allocation when processing high-resolution images. To overcome these challenges, we propose an adaptive object detection method on rainy days. This approach incorporates a rain removal model based on Principal Component Analysis (PCA) as a preprocessing step and combines it with a target detector to improve the image recognition effectiveness. It adaptively predicts the hyperparameters for the Preprocessing Refinement Module (PRM) by inputting low-resolution images into Convolutional Neural Networks (CNNs). This step not only eliminates the need for manual intervention but also reduces the computational complexity of the CNN training step. Then, the PRM uses these predicted hyperparameters for rain-removal and image enhancement. Thus, APCA-Net can improve the precision of object detection. Our experimental results on VOC-based synthetic datasets confirm the effectiveness of this method.