A transfer-learning TNet for multi-task railroad inspection

This paper introduces a novel transfer learning framework designed to simultaneously perform object detection and semantic segmentation, with specific application to railroad inspection. The proposed approach consists of two distinct yet integrated components: a frozen, pre-trained module and a newly developed trainable module named TNet. The frozen component leverages a pre-trained model initially trained to detect and localize rail track components such as clips and spikes, generating bounding boxes around both sides of the rails. The TNet module, on the other hand, introduces a set of innovative structures specifically aimed at identifying and segmenting rail surface defects with high accuracy. The novelty of TNet lies in the introduction of new components, a Bias Correction Network (BCN), a Channel-Score Attention (CSA) module, and a single-step Upsample Block and their coordinated integration to improve transfer learning performance. The BCN is designed to mitigate biases inherent in the feature maps from the frozen pre-trained network, ensuring more balanced and accurate representations tailored explicitly for defect segmentation. The CSA module optimizes the representation of multi-channel feature maps by assigning adaptive importance scores, thereby reducing redundancy and computational complexity. Furthermore, the single-step Upsample Block enhances the segmentation accuracy by efficiently reconstructing high-resolution feature maps through combined pixel shuffle and transposed convolution operations. The entire TNet remains exceptionally lightweight, with a total size under 5 MB and a computational cost of merely 0.96 GFLOPs. Accordingly, the inference speed of our model remains comparable to running the upstream detection task alone. Experimental evaluation demonstrates that the proposed model achieves a pixel accuracy of 93.20 % for defect segmentation, significantly surpassing existing benchmarks. These results underscore TNet's potential as a robust, real-time solution for railroad safety and maintenance inspections, with broader applicability to other multi-task computer vision challenges.