<![CDATA[This study develops an image classification system for chili pepper seeds using a hybrid approach that combines a manual Convolutional Neural Network (CNN) based on Sobel filters and Learning Vector Quantization (LVQ). The dataset consists of 900 images from three chili pepper varieties, namely gendot, raksasa, and rawit, with 300 images for each class captured using a smartphone camera. Unlike conventional CNN architectures that rely on trainable convolutional kernels, this study adopts predefined Sobel filters as fixed feature extractors combined with Learning Vector Quantization as the classifier, resulting in a lightweight classification framework that does not require a training process in the feature extraction stage. The CNN architecture employs three convolution-pooling layers (depth-3), producing feature maps of size 16×16, which are then transformed into 256-dimensional feature vectors through a flattening process. LVQ training was conducted for 30 epochs with an initial learning rate of 0.001 and five prototypes assigned to each class. The model was evaluated using a 5-fold cross-validation scheme and obtained a mean accuracy of 80.78% with a standard deviation of 4.07% across folds. Based on the combined confusion matrix from all five folds, the gendot class obtained a precision of 1.00 and a recall of 0.93, the raksasa class obtained a precision of 0.75, a recall of 0.69, and an F1-score of 0.72, and the rawit class obtained a precision of 0.71, a recall of 0.80, and an F1-score of 0.75, with most misclassifications occurring between the raksasa and rawit classes due to their morphological similarity. These findings show that the hybrid CNN–LVQ approach can be applied to chili seed classification without requiring specialized hardware or a pre-trained model.]]>
