Image Processing Based on Support Vector Machines

This paper investigates image classification implementation using Libsvm, focusing specifically on object classification with five fruit categories as research subjects. The comprehensive workflow for image classification involves the following key stages:

1. Image Sample Collection

- Acquisition of substantial fruit images from web sources, including five categories: apples, bananas, oranges, lemons, and kiwifruits.

- Image filtering and deduplication to ensure balanced quantity distribution across fruit categories.

2. Image Preprocessing

- Image cropping and resizing to uniform dimensions (e.g., 256x256 pixels) using OpenCV's resize() function for consistent processing.

- Image enhancement techniques including histogram equalization for brightness adjustment, Gaussian filtering for noise reduction, and sharpening kernels for edge enhancement.

3. Feature Vector Extraction

- Extraction of multidimensional feature vectors incorporating color histograms in HSV space, texture features using Local Binary Patterns (LBP), and shape descriptors through Hu moments.

- Feature normalization using Z-score standardization or min-max scaling to ensure all features operate on comparable scales.

4. Libsvm-based Model Training

- Implementation of classifier training using Libsvm's svm_train() function with k-fold cross-validation for optimizing parameters (C, gamma) via grid search.

- Determination of the final classifier model through performance evaluation metrics including accuracy and F1-score.

5. Classification Testing on New Images

- Classification of unseen test images using svm_predict() function to evaluate model accuracy and generalization capability.

- Experimental result analysis with confusion matrix visualization and proposal of optimization strategies like feature selection or kernel function adjustment.

Through this detailed workflow, the study provides comprehensive insights into Libsvm-based image classification implementation, offering valuable references and methodologies for related research domains.