TINYCD: A (Not So) Deep Learning Model For Change Detection (2207.13159v2)

Abstract: In this paper, we present a lightweight and effective change detection model, called TinyCD. This model has been designed to be faster and smaller than current state-of-the-art change detection models due to industrial needs. Despite being from 13 to 140 times smaller than the compared change detection models, and exposing at least a third of the computational complexity, our model outperforms the current state-of-the-art models by at least $1\%$ on both F1 score and IoU on the LEVIR-CD dataset, and more than $8\%$ on the WHU-CD dataset. To reach these results, TinyCD uses a Siamese U-Net architecture exploiting low-level features in a globally temporal and locally spatial way. In addition, it adopts a new strategy to mix features in the space-time domain both to merge the embeddings obtained from the Siamese backbones, and, coupled with an MLP block, it forms a novel space-semantic attention mechanism, the Mix and Attention Mask Block (MAMB). Source code, models and results are available here: https://github.com/AndreaCodegoni/Tiny_model_4_CD

• The paper introduces TinyCD, a compact change detection model that uses low-level features to achieve superior performance on remote sensing datasets.
• It employs a Siamese U-Net framework with a novel Mix and Attention Mask Block to effectively merge spatial and temporal information.
• Experimental results show TinyCD reduces model complexity while improving F1 and IoU scores, outperforming more complex state-of-the-art models.

Overview of "TinyCD: a (not so) Deep Learning Model for Change Detection"

The paper "TinyCD: a (not so) Deep Learning Model for Change Detection" presents a novel approach to change detection (CD) tasks in the remote sensing domain using a deep learning framework. The model, named TinyCD, is notable for its significantly reduced size and complexity compared to existing state-of-the-art (SOTA) models. Despite being substantially smaller—by factors ranging from 13x to 140x—TinyCD achieves superior performance, improving key metrics such as F1 score and Intersection over Union (IoU) by at least 1% on the LEVIR-CD dataset and over 8% on the WHU-CD dataset.

Key Design Elements

The architecture of TinyCD employs a Siamese U-Net framework that takes advantage of low-level feature abstraction. It uses both spatial and temporal information to identify changes between co-registered images captured at different times. A notable feature of TinyCD is the Mix and Attention Mask Block (MAMB), which incorporates a simple mixing strategy combined with a multi-layer perceptron (MLP) to perform space-semantic attention operations. The emphasis on effective feature utilization without heavy reliance on parameter-rich backbones allows for a lightweight yet effective model.

Contributions

The paper highlights several critical contributions to the field:

• Use of Low-Level Features: TinyCD successfully demonstrates that low-level features are sufficiently expressive for CD tasks when utilized effectively. This insight leads to a significant reduction in model size and computational requirements without sacrificing accuracy.
• Novel Mixing Strategy: The introduction of MAMB provides a robust mechanism to merge features from two temporal images, which improves the representation of spatial-temporal correlations.
• Efficient Attention Mechanism: The integration of a fast attention mechanism allows the model to refine prediction masks during the up-sampling phase, further enhancing output resolution.
• Lightweight Design: TinyCD manages to cap the number of model parameters at approximately 300,000, a significant reduction from typical deep networks, making it feasible for real-time applications and deployment on constrained hardware resources.

Experimental Results

In experiments, TinyCD outperforms several prominent CD models, including those incorporating more complex architectures like ResNet and Vision Transformers. On the LEVIR-CD dataset, TinyCD achieves an F1 score of 91.05% and IoU of 83.57%, while on the WHU-CD dataset, it scores 91.74% F1 and 84.74% IoU. These results not only demonstrate the model's high effectiveness but also its efficiency, evidenced by its minimal computational complexity of just 1.45 GFLOPs.

Implications and Future Directions

The implications of this work are significant for applied machine learning in environments where computational resources are limited. By demonstrating that effective topological and domain-specific changes can be detected using compact models, this research opens avenues for deploying advanced computer vision models in real-world, resource-constrained scenarios, such as smart city infrastructure and remote sensing for environmental monitoring.

Looking forward, the paper suggests that future research may involve exploring hybrid architectures that incorporate both local and global feature relationships, perhaps leveraging the complementary strengths of CNNs and Transformer-based models. Additionally, application domains beyond remote sensing, such as industrial automation and surveillance, might benefit from similar lightweight architectures.

In conclusion, TinyCD represents an impressive stride in the development of efficient CNNs for change detection, laying the groundwork for further exploration into minimizing architectural complexity while maximizing model efficacy.