Optimizing image segmentation processes using deep learning: a neural network-based approach

Image segmentation is a key step in many computer vision systems, particularly in medical imaging, remote sensing, robotics, and precision agriculture. Over the years, a wide range of traditional approaches have been developed for image segmentation, such as binary thresholding (Otsu's method), the watershed method, and K-means clustering, which leverage domain-specific knowledge to efficiently solve segmentation problems in particular application contexts. Unfortunately, however, these methods show significant limitations when faced with the increasing complexity of visual scenes and data heterogeneity. In reality, the fundamental principle of image segmentation is to classify each pixel based on similar attributes. The goal of this pixel-by-pixel classification is to achieve a process that is global, relatively insensitive to noise, locally accurate, and incorporates multi-scale filtering mechanisms. However, traditional methods cannot simultaneously meet these requirements. Faced with the limitations of traditional methods, the emergence of deep learning approaches, particularly those based on neural networks and especially convolutional neural networks, has established itself as an effective solution for image segmentation. This article proposes an approach to optimizing segmentation processes based on the comparative analysis and improvement of several deep neural architectures. The study evaluates the performance of models such as the linear network, Simple U-Net, Extended U-Net, U-Net, DeepLabV3+, and SegFormer, incorporating optimization strategies including data augmentation, fine-tuning, the addition of attentional mechanisms, and the proposal of a specific metric to track performance changes during neural network model training. The experimental results demonstrate a significant improvement in segmentation accuracy, measured by the IoU, Dice Score, and F1-score indicators, confirming the effectiveness of deep neural networks optimized for complex visual environments.