Photovoltaic Panel Defect Detection via Multi-scale Siamese Convolutional Fusion Network with Information Bottleneck Theory

This paper presents a solution to the challenges in detecting rare faults in photovoltaic panels, where sample imbalance and diverse damage types lead to a wide range of failure types. A multi-scale Siamese convolutional neural network (MsSCNN) with an information bottleneck is developed to address sample imbalance by sharing parameters, which can accurately distinguish various fault types by effectively capturing similarities and differences between faults. First, a multi-scale convolutional fusion module is embedded in the MsSCNN for capturing detailed information at multiple levels to improve the detection of complex defects. Additionally, a depth-wise separable convolution module is further employed to increase the expression of different domain features without adding complexity to the model. However, the comprehensive extraction of detailed information can result in redundancy in the feature space. To mitigate this, the information bottleneck method formed by a mutual information loss function and bottleneck module is introduced into the backbone network to compress information and reduce dimensionality ensuring the retention of key features from high-dimensional input data. Experimental results on different dataset demonstrate the superiority of our method over multiple existing deep Siamese network approaches.