芝麻模型训练样本数据
收藏资源简介：
芝麻模型训练样本数据1、数据采集：利用大疆御3M无人机，利用自动拼图得到1cm-3cm分辨率的无人机正射影像数据，并设置CGCS2000 / 3°投影坐标系，影像分辨率以及影像坐标系等参数同步加入至影像数据中。 2、数据预处理以及数据标注：首先，选择适用于样本的影像，明确裁切区域并绘制范围矢量；利用矢量数据裁切影像，采用在线标记点位并添加作物属性的方式，以点位为中心点进行裁剪，生成指定尺寸（如256*256）的影像切片。 3、数据集设置以及模型选择：按照7:2:1的比例设置训练集、验证集和测试集。使用自行搭建的TransCNN-Vision模型进行训练。 5、训练设置：模型选择与初始化以vision_transformer的large模型为权重文件，初始化常规模型参数后，读取数据集文件夹个数确定模型分类数，最优学习率为0.001，同时冻结出head和pre_logits外的所有权重参数，batchsize，根据当前设备现存余量自动调整，默认值为16，根据样本分辨率动态调整patch_size用于提高不同分辨率下的特征捕获能力。最后利用自适应高精度模型保存策略，自动保存训练精度mDice（Mean Dice Coefficient）指标最高模型的模型作为最佳模型。mDice计算公式如下： mDice=2*|X∩Y|/(|X|+|Y|)。 训练mDice指数基于对测试集样本数据预测并计算获得，其中|X∩Y|为预测结果与真实标注的交集，| X |和| Y |分别为预测结果与真实标签的各自的数量之和。 5、模型精度评估：通过在真实影像中进行模型提取并人工校正，实现对模型在真实场景中的提取效果。利用提取错误率和提取遗漏率指标来评估被识别物模型的提取能力，提取错误率用于评估模型提取结果中不是合理的比例，提取遗漏率用于评估模型提取结果依然没有提取出被识别物的比例，提取遗漏率越接近4.2%，提取错误率越接近6.2%，表明当前提取结果准确率越高，能够降低的成本越高。提取错误率计算公式如下：（|X|-|X∩Y|）/|X|。 提取遗漏率计算公式如下：（|Y|-|X∩Y|）/|Y|。其中，|X∩Y|为正确识别为被识别物的数量，| X |和| Y |分别为预测和真实的被识别数量。考虑到真实场景的复杂性，提取错误率和提取遗漏率保持在10%以内即可视为结果具有较高的准确性。
1. Data Collection: Utilize the DJI Mavic 3M UAV to obtain 1cm–3cm resolution orthophoto imagery via automatic mosaicking. Adopt the CGCS2000 / 3° projection coordinate system, and synchronously embed parameters such as image resolution and coordinate system into the image data. 2. Data Preprocessing and Annotation: First, select eligible images, define cropping regions and draw vector boundaries; crop the images using the vector data. Then, mark points online and add crop attributes, take each marked point as the center to generate image slices with a specified size (e.g., 256×256). 3. Dataset Split and Model Selection: Split the dataset into training, validation and test sets at a ratio of 7:2:1. Use a self-developed TransCNN-Vision model for training. 5. Training Configuration: Initialize the model with the weights of the vision transformer large model. After initializing the conventional model parameters, determine the number of model classification categories by reading the number of dataset folders. The optimal learning rate is set to 0.001. Freeze all weight parameters except the head and pre_logits layers. The batch size is automatically adjusted based on the available memory of the current device, with a default value of 16. Dynamically adjust the patch_size according to the sample resolution to enhance feature capture capability across different resolutions. Finally, adopt an adaptive high-precision model saving strategy to automatically save the model with the highest training mDice (Mean Dice Coefficient) metric as the best model. The formula for mDice is as follows: mDice = 2×|X∩Y|/(|X|+|Y|). The training mDice metric is calculated based on predictions of test set samples, where |X∩Y| represents the intersection of prediction results and ground-truth annotations, and |X| and |Y| represent the total number of instances in the prediction results and ground-truth labels, respectively. 5. Model Accuracy Evaluation: Evaluate the model's extraction performance in real-world scenarios by extracting results from real images and performing manual correction. Use the extraction error rate and extraction omission rate metrics to assess the model's ability to identify target objects. The extraction error rate is used to evaluate the proportion of unreasonable results in the model's extraction outputs, with the formula: (|X| - |X∩Y|)/|X|. The extraction omission rate is used to evaluate the proportion of target objects that are not extracted in the model's results, with the formula: (|Y| - |X∩Y|)/|Y|. Here, |X∩Y| represents the number of correctly identified target objects, while |X| and |Y| represent the total number of predicted and ground-truth target objects, respectively. It is noted that the closer the extraction omission rate is to 4.2% and the extraction error rate is to 6.2%, the higher the accuracy of the current extraction results and the greater the cost reduction that can be achieved. Considering the complexity of real-world scenarios, keeping both the extraction error rate and extraction omission rate within 10% is considered to indicate high accuracy of the results.
