基于改进DeepLabV3+的青贮饲料收获机跟随料车车斗识别方法

Recognition Method of Silage Harvester Follow-up Trailer Hopper Based on Improved DeepLabV3+

摘要:

青贮饲料收获机跟随料车车斗的准确识别是实现自动抛料装车的首要前提，针对青贮收获田间作业环境复杂、料车种类繁多，现有料车车斗识别方法难以适用各类装载料车识别的问题，利用图像语义分割技术提出一种改进DeepLabV3+的青贮饲料收获机跟随料车车斗识别方法。该方法将轻量化的MobileNetV3网络作为主干网络进行特征提取，降低模型复杂度;在ASPP模块中以条形池化替换原有的平均池化加强对车斗的识别能力，并将原有的3个空洞卷积分支扩充为4个空洞深度可分离卷积分支，提升对大目标的识别能力;在解码器部分融合了多尺度特征，最大程度地弥补损失的车斗边界细节特征，并在拼接融合后添加C2f-GS模块进一步增强模型的特征融合能力。在构建的料车车斗数据集上设计消融试验和对比试验对模型改进的有效性进行验证，试验结果表明，改进DeepLabV3+模型mIoU和mPA达到94.25%和96.62%，分割效果优于其他分割模型，且模型参数量和浮点运算量显著低于其他模型。与原DeepLabV3+模型相比，改进模型的mIoU和mPA分别提高4.77、3.71个百分点，识别准确性得到有效提升。利用基于RANSAC算法的料车车斗拟合方法对分割后的结果进行车斗边界拟合，对不同情况下的分割结果进行拟合试验，结果表明拟合的车斗与原始标注车斗的mIoU平均值达到94.49%，二者质心点位置偏差仅为25.47像素，拟合结果基本能够覆盖车斗区域，满足青贮饲料收获机田间作业情况下的需求。

Abstract:

Accurate recognition of the silage harvester follow-up trailer hopper is the first prerequisite for realizing automatic throwing and loading, for the problem that the complex operational environment of silage harvesting, the variety of trailer types, and existing methods for trailer hopper recognition struggle to apply to all kinds of loading hopper identification, using the technology of image semantic segmentation, a silage harvester follow-up trailer hopper recognition method was proposed with improved DeepLabV3+. The method took the lightweight MobileNetV3 network as the backbone network for feature extraction to reduce the model complexity. In the ASPP module, the original average pooling was replaced by strip pooling to enhance the recognition ability of the hopper, and the original three atrous convolution branches were expanded into four atrous depthwise separable convolution branches to improve the recognition of large targets. Moreover, the multi-scale features in the decoder part were integrated to compensate for the lost hopper boundary detail features. After concatenation and fusion, the C2f-GS module was added to further enhance model's feature fusion capabilities. The ablation tests and comparison tests were conducted on the constructed hopper dataset to verify the effectiveness of the improved model, and the test results showed that the values of the mean intersection over union (mIoU) and mean pixel accuracy (mPA) of the proposed improved DeepLabV3+ model reached 94.25% and 96.62%, which was better than that of the other segmentation models, and the model parameters and FLOPs were significantly lower than that of the other models. Compared with the original DeepLabV3+ model, the mIoU and mPA of the improved model were increased by 4.77 and 3.71 percentage point, model recognition accuracy was effectively enhanced. Using the hopper fitting method based on the RANSAC algorithm, the boundary of the hopper was fitted from the segmented results, fitting experiments conducted under various conditions revealed that the mIoU of the fitted hopper and the original labeled image reached 94.49%, and the deviation of the center-of-mass point position between the two was only 25.47 pixels, which indicated that the fitting result basically could cover the area of the hopper , meeting the operational requirements for silage harvesting in the field.