Most neural networks recognize objects based on their contours, which means that their accuracy is effectively independent of the colour temperature of the light that illuminates the test object. Determining factors in the recognition of wood species are not its contours but rather the surface structure and texture. Hence, the accuracy of standard neural networks in the recognition of wood species depends on the colour temperature of the light. The aim of this study is to develop a neural network for the recognition of selected wood species regardless of the colour temperature of light. A total of 52 neural networks were created using MATLAB 2019a software, including three layers: an input, hidden (with 10, 20, 50 and 100 neurons) and output layer. Neural networks were trained, validated and tested using photographs of beech, larch, spruce and pine taken at colour temperatures of 2700, 4000 and 6500 K and additionally tested using photographs taken at colour temperatures of 3500 and 5500 K. The neural networks were trained using coloured and grey scale images (adjusted with averaging and/or by emboss or sharpen kernels). During the additional test, the highest accuracy (97.9%) was observed in the neural network trained with grey scale images adjusted with averaging and emboss kernels. The algorithm that recognized the wood species based on the identical classification of at least 3 out of 5 photographs from different areas of the same sample was even more accurate (99.99%).

大多数神经网络根据物体的轮廓识别物体，这意味着它们的准确性实际上与照亮测试物体的光的色温无关。识别木材种类的决定因素不是其轮廓，而是表面结构和纹理。因此，标准神经网络识别木材种类的准确性取决于光的色温。本研究的目的是开发一种神经网络，无论光的色温如何，都可以识别选定的木材种类。使用 MATLAB 2019a 软件共创建了 52 个神经网络，包括三层：输入层、隐藏层（具有 10、20、50 和 100 个神经元）和输出层。神经网络使用山毛榉、落叶松、云杉和松树在 2700、4000 和 6500 K 色温下拍摄，并使用在 3500 和 5500 K 色温下拍摄的照片进行额外测试。神经网络使用彩色和灰度图像（通过平均和/或浮雕或浮雕调整）进行训练。锐化内核）。在附加测试期间，在使用平均和浮雕内核调整的灰度图像训练的神经网络中观察到最高准确度 (97.9%)。根据来自同一样本不同区域的 5 张照片中至少 3 张的相同分类来识别木材种类的算法甚至更准确 (99.99%)。神经网络使用彩色和灰度图像（通过平均和/或浮雕或锐化内核进行调整）进行训练。在附加测试期间，在使用平均和浮雕内核调整的灰度图像训练的神经网络中观察到最高准确度 (97.9%)。根据来自同一样本不同区域的 5 张照片中至少 3 张的相同分类来识别木材种类的算法甚至更准确 (99.99%)。神经网络使用彩色和灰度图像（通过平均和/或浮雕或锐化内核进行调整）进行训练。在附加测试期间，在使用平均和浮雕内核调整的灰度图像训练的神经网络中观察到最高准确度 (97.9%)。根据来自同一样本不同区域的 5 张照片中至少 3 张的相同分类来识别木材种类的算法甚至更准确 (99.99%)。