L’apprentissage profond pour la classification d’images Agricoles

Abstract

Convolutional neural networks (CNNs) have gained significant popularity in image classification tasks. However, designing an optimal CNN model can be challenging due to the vast space of possible combinations of layer numbers and associated hyperparameter values. Selecting the best CNN model for a specific task often requires extensive training of numerous models, resulting in time-consuming processes. To address this problem, we propose a novel automated approach for CNN architecture design. Our proposed framework consists of two alternative variants : the first one utilizes pre-trained models, while the second one utilizes a bloc system as a backbone for the two variants, we employ two different evolutionary algorithms, namely gray wolf optimization and genetic algorithm. The primary objective of our research is to automatically generate and evaluate candidate CNN architectures for plant seedling classification, specifically distinguishing between weed and crop seedlings. In addition, we introduce an encoding system for representing CNN architectures and their corresponding hyperparameters. By combining evolutionary algorithms with transfer learning and the bloc system, we aim to extract meaningful features from images. Through extensive experimentation, our proposed method achieves exceptional accuracy, surpassing state-of-theart approaches, with a validation accuracy of up to 97.83%. This approach provides a revolutionary tool for enhancing the accuracy of CNN models, tailored specifically to the dataset under consideration.