Scribd
Upload
10 views20 pages

Robot

Robot

Uploaded by

Shafaque Malik
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
10 views20 pages

Robot

Robot

Uploaded by

Shafaque Malik
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 20

CONVOLUTIONAL AUTOENCODER FOR

HAZE REMOVAL FOR MULTI IMAGES


ALONG WITH DIFFERENT IMAGE TYPES

Student Names: Tariq Ali, Hafiz Eman Haider Butt, Iftekhar Ahmed
Supervisor: Mis Shafaque Saira Malik
Introduction / Background

• Autoencoders are a powerful type of neural network that excel in unsupervised


learning by encoding input data into a compressed form and then reconstructing
the original data from this compressed representation.

• When applied to image dehazing, autoencoders can effectively learn to improve


the quality of images obscured by atmospheric conditions such as fog or mist.

• Atmospheric haze is a common phenomenon that significantly degrades the


quality of images captured in outdoor environments. It reduces visibility and
impairs the clarity of visual content by scattering light, leading to blurred images
with diminished contrast and detail.
Objectives of the Study

• Model that significantly improves the visual quality of hazy images

• Utilize the feature extraction strengths of CNNs (traditional methods may fail to
address.)

• Implement an autoencoder architecture to learn and reconstruct haze-free images


from hazy inputs.

• maintaining the integrity of the visual information.


Research Question

• RQ1. How can we adjust the CNN layers in an autoencoder to improve image
dehazing?

• RQ2. What is the impact of using simpler CNN architectures on the quality of
dehazed images?
Limitations of the Study

• CNN-based autoencoders require large amounts of high-quality paired data for


effective training.

• It involves significant computational resources, including powerful GPUs.

• This can be a barrier for organizations with limited hardware capabilities.

• While CNNs can learn to handle specific types of haze.


• They may struggle with new or extreme haze conditions not represented in the
training data.

• The process of removing haze and reconstructing clear images can sometimes
introduce artifacts or distortions.
Variable of Study

• Number of layers in the CNN-based autoencoder.

• Types of convolutional layers used (e.g., standard, dilated).

• Activation functions applied (e.g., ReLU, Leaky ReLU).

• Pooling methods employed (e.g., max pooling, average pooling).

• Number of filters or channels in each layer of the network.


Hypothesis

• H0:
Convolutional Neural Networks, with their hierarchical feature extraction capabilities,
can effectively learn and represent complex haze patterns present in images. This advanced
feature learning will enable the autoencoder to more accurately capture and compensate for
haze effects, resulting in clearer and more detailed dehazed images.

• H1:
The autoencoder's architecture, comprising an encoder and decoder network, will
facilitate improved reconstruction of haze-free images from hazy inputs. The network's
ability to compress and reconstruct image data will lead to better preservation of image
details and contrast compared to methods that rely on handcrafted features or simpler
algorithms.
Research Methodology
Research Analysis Tools

• Mean Squared Error (MSE)


• Peak Signal-to-Noise Ratio (PSNR)
• Structural Similarity Index (SSIM)
• Matplotlib
• TensorFlow
• OpenCV
• Plot
• Scikit-learn
• Keras (for TensorFlow)
• PyTorch
Data Collection and variable of the Study

• Hazy-Clear Image Pairs


• Dataset Examples
• Data Preprocessing
• Data Splitting
• Input Variables
• Output Variables
• Performance Metrics
• Training Parameters
• Model Architecture Variables
• Loss Function Components
Sampling

Batch Sampling
• Number of images processed together in one forward/backward pass. Common
values are 16, 32, or 64.
• Shuffle the dataset to ensure a diverse batch for each iteration.

Learning Rate Scheduling


• Start with a value like 0.001.
• Use learning rate scheduling to decrease the rate as training progresses.
Results and Discussion
PSNR (Peak Signal-to-Noise Ratio)
• Higher values indicate better quality of the generated images.
SSIM (Structural Similarity Index)
• Measures the perceived quality of images. Values closer to 1 indicate better
quality.
MSE (Mean Squared Error)
• Lower values represent better performance with less error.
RMSE (Root Mean Squared Error)
• Provides a measure of the average magnitude of the errors, with lower values
indicating better performance.
MAE (Mean Absolute Error)
• Average magnitude of errors in predictions, where lower values are preferable.
Inference Time
• Time taken to make a prediction on a single image. Lower values are preferable
for real-time applications.
Training Time
• Total time required to train the model on the dataset. Shorter times are generally
better, but this may vary based on model complexity and dataset size.
Number of Parameters
• Total number of trainable parameters in the model. Models with fewer parameters
are generally more efficient but might be less capable.
Image Resolution
• The size of images used in training and inference. Higher resolutions typically
allow for better image quality.
Conclusion and Future Work

Conclusion
The use of autoencoders with Convolutional Neural Networks (CNNs) for image
dehazing has shown notable success in enhancing image quality by effectively
removing haze and restoring clarity.
These models leverage CNNs' powerful feature extraction capabilities to
reconstruct haze-free images from hazy inputs, demonstrating significant
improvements in visual quality and detail preservation.
The approach is versatile and adaptable, handling various haze intensities and
conditions, making it suitable for diverse applications.
Future Work
Future research should focus on several key areas to further advance the field.
Exploring more advanced network architectures, such as U-Net or GAN-based
models.
Expanding and diversifying datasets to include a wider range of real-world hazy
images will improve model robustness and generalization.
Addressing the challenge of synthetic-to-real data transfer is also crucial for
practical deployment.
Developing user-centric approaches and interactive tools that allow for
customizable dehazing settings based on individual needs will enhance practical
utility.
References

• Xu, P., Zhang, Z., Mo, X., & Sun, F. (2018). Deep haze removal with multi-scale
fusion network. Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2018, 7473-7481.
• Shen, C., Zhang, T., & Reid, I. (2019). Single image dehazing using multi-scale
convolutional neural network. IEEE Transactions on Image Processing, 28(12),
5710-5722.
• Zhang, Y., Zhang, Y., & Chuang, Y.-Y. (2017). Autoencoder-based convolutional
neural network for image dehazing. Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 2017, 3833-3841.
• Dong, B., Yu, X., & Qiao, Y. (2018). DehazeNet: An end-to-end system for single
image haze removal. IEEE Transactions on Image Processing, 27(12), 5982-5995.
• Hu, X., Li, H., & Xu, Z. (2019). AOD-Net: All-in-one dehazing network. IEEE
Transactions on Image Processing, 28(6), 2618-2629.
• Li, H., Hu, X., & Wang, Z. (2018). Dense-CRF dehazing network for single image
haze removal. Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2018, 7642-7650.
Thank You

You might also like