Date Palm data.

Agriculture has become an essential field of study and is considered a challenge for many researchers in computer vision specialization. The early detection and classification of plant diseases are crucial for preventing growing diseases and hence yield reduction. Although many state-of-the-artwork proposed various classification techniques for plant diseases, still face many challenges such as noise reduction, extracting the relevant features, and excluding the redundant ones. Recently, deep learning models are noticeable as hot research and are widely used for plant leaf disease classification. Although the achievement with these models is notable, still the need for efficient, fast-trained, and few-parameters models without compromising on performance is inevitable. In this work, two approaches of deep learning have been proposed for Palm leaf disease classification: Residual Network (ResNet) and transfer learning of Inception ResNet. The models make it possible to train up to hundreds of layers and achieve superior performance. Considering the merit of their effective representation ability, the performance of image classification using ResNet has been boosted, such as diseases of plant leaves classification. In both approaches, problems such as variation of luminance and background, different scales of images, and inter-class similarity have been treated. Date Palm dataset having 2631 colored images with varied sizes was used to train and test the models. Using some well-known metrics, the proposed models outperformed many of the recent research in the field in original and augmented datasets and achieved an accuracy of 99.62% and 100% respectively.农业是至关重要的研究领域，同时也为计算机视觉专业的诸多研究者带来了挑战。植物病害的早期检测与分类，对于遏制病害扩散、避免作物减产具有关键意义。尽管诸多前沿研究已针对植物病害提出了多样的分类技术，但该领域仍面临诸多挑战，例如降噪、提取有效特征并剔除冗余特征等。近年来，深度学习模型作为研究热点备受关注，被广泛应用于植物叶片病害分类任务中。尽管此类模型已取得显著成果，但学界仍亟需在不牺牲模型性能的前提下，开发高效、快速训练且参数量较少的病害分类模型。本研究针对椰枣叶片病害分类任务，提出了两种深度学习方案：残差网络（Residual Network）以及基于Inception ResNet的迁移学习方法。此类模型支持对数百层网络进行训练，并可获得优异的性能表现。鉴于其出色的特征表征能力，基于ResNet的图像分类任务性能得到了显著提升，在植物叶片病害分类领域亦是如此。针对两种方案，本研究均解决了亮度与背景差异、图像尺度不一以及类间相似性等常见问题。本研究采用椰枣数据集（Date Palm Dataset）开展模型训练与测试工作，该数据集包含2631张尺寸各异的彩色图像。通过多项经典评价指标验证，所提模型在原始数据集与增强数据集上的表现均优于该领域诸多近期研究，分别实现了99.62%与100%的分类准确率。