Character Recognition based on Convolutional Neural Network and Hough Transform

What is it about?

This study explores the integration of the Hough Transform (HT) with Convolutional Neural Networks (CNNs) to improve pattern recognition accuracy in image collections, specifically focusing on detecting straight lines. The Hough Transform, a classical image processing technique, is used to preprocess images by enhancing geometric features, while CNNs leverage deep learning to classify these features effectively. Methodology: The proposed approach combines the Extended Standard Hough Transform (ESHT) with a CNN model. The ESHT preprocesses images to highlight straight lines, which are then fed into a CNN architecture consisting of three convolutional layers, each followed by a Max Pooling layer, and a fully connected layer for classification. The model uses ReLU activation functions and the Adam optimizer for training. Data augmentation techniques were applied to a small, custom dataset of 32x32 images to expand the training set to 10,200 images, split into 70% training, 15% testing, and 15% validation subsets. Results: Experimental results demonstrated that the integration of ESHT with CNNs improved classification accuracy. The model achieved 70% accuracy with ESHT preprocessing, compared to 66.67% without it. The confusion matrix, precision, recall, and F-measure metrics confirmed the enhanced performance, particularly in reducing false positives and negatives. Visual comparisons (Figures 3 and 4) showed clearer detection of lines in ESHT-processed images. Challenges and Future Work: Limitations included a small, simulated dataset and hardware constraints (e.g., 32x32 image size due to memory limitations). Future directions include diversifying datasets, comparing the model with other neural networks, and exploring applications in fields like transportation. Conclusion: The study validates the hypothesis that combining HT with CNNs enhances pattern recognition accuracy, offering a promising hybrid approach for computer vision tasks. Further improvements could be achieved with larger datasets and more robust computational resources. Character_recognition_based_on_Convolutional_Neural_Network_and_Hough_Transform.pdf PDF 396.02KB Je veux un résumé en anglais d'une page. Summary: Character Recognition Based on Convolutional Neural Network and Hough Transform This study presents a hybrid approach combining Convolutional Neural Networks (CNNs) with the Hough Transform (HT) to enhance handwritten character recognition, achieving an accuracy of 98.53% on Kaggle’s A-Z Handwritten Dataset. The integration of HT addresses challenges like character slant and size variations, while CNNs leverage deep learning for robust feature extraction and classification. Methodology Preprocessing: Binarization: Converts grayscale images to binary to separate text from background. Smoothing: Reduces noise and artifacts. Normalization: Standardizes character sizes to minimize distortion. Skeletonization: Simplifies characters to one-dimensional structures for easier processing. Hough Transform (HT): Linear HT (LHT): Detects straight-line features in characters (e.g., ‘W’, ‘N’). Circular HT (CHT): Identifies curved features (e.g., ‘O’, ‘Q’). HT enhances feature extraction by isolating geometric primitives (lines, curves) before CNN processing. CNN Architecture: Three convolutional layers with ReLU activation and max-pooling for dimensionality reduction. A fully connected layer for classification using softmax. Trained on 372,450 images (28x28 pixels), augmented to handle variability.

Featured Image

Photo by Jr Korpa on Unsplash

Photo by Jr Korpa on Unsplash

Why is it important?

Key Results Without HT: Misclassification of similar characters (e.g., ‘N’ as ‘M’, ‘M’ as ‘W’). With HT: Correct recognition of all tested characters (Fig. 5). Comparative Performance: Outperformed other methods (Table 2), achieving 98.53% accuracy vs. 95.36% (EMNIST) and 94% (custom datasets).

Perspectives