Consistency Guided Network for Degraded Image Classification

This paper presents a novel end-to-end Consistency Guided Network (CG-Net) for degraded image classification. CG-Net effectively tackles the challenges arising from the inconsistency of category and feature distribution between degraded and clear images.

A. Overview of the Network Architecture

CG-Net employs a teacher-student model, comprising a Source Network (S-Net) trained on clear images and a Target Network (T-Net) trained on degraded images. During training, the S-Net provides valuable hints to the T-Net. This architecture allows CG-Net to fully leverage the relevant information related to image classification between clear and degraded images.

To address the issue of category distribution inconsistency, we introduce a Category Consistency Loss (CCL) (Section III-C). CCL guides the model towards learning a category distribution that aligns more closely with the distribution of clear images.

To address the feature distribution inconsistency, we propose a Semantic Consistency Loss (SCL) (Section III-D). SCL enforces the model to learn a more robust feature representation that is more consistent with the feature representation of clear images.

Furthermore, to handle the misalignment of visual attention heatmaps between degraded and clear images, we present a Visual Attention Alignment Loss (VAAL) (Section III-E). VAAL guides the network to focus on semantically informative regions that are aligned with those identified in clear images.

Importantly, our approach is versatile and suitable for various types of degraded images. Initially, S-Net and T-Net are trained independently on clear and degraded images, respectively. Subsequently, we extract category distribution, semantic feature distribution, and visual attention from both clear and degraded images. We then calculate CCL, SCL, and VAAL based on these extracted features. Finally, we fix S-Net and optimize T-Net by simultaneously minimizing classification loss, CCL, SCL, and VAAL.

The baseline network used in our implementation is ResNet-50 [5]. A detailed structure of the network is provided in Table I, where 'C' represents the total number of classes in the dataset.

Complexity Analysis

While the parameter count of CG-Net is twice that of the baseline network, the training process for degraded images converges rapidly. For testing, we utilize only the trained target network. Therefore, the parameters and running time of our proposed method are essentially equivalent to those of the baseline, making our network computationally efficient.

Consistency Guided Network for Degraded Image Classification