Abstract: To solve problems of background interference and information loss of ripe strawberries under natural state, a method for strawberry recognition was proposed based on deep residual learning in this study. Firstly, the depthwise separable convolution was introduced to reduce the residual network parameters, the features of ripe strawberries were extracted from different aspects, and the strawberries in the classification layer were identified through the crossentropy loss function. Additionally, the feature weights were learned by embedding compression and excitation modules, and feature recalibration was used to improve the learning and representation properties of the network. Finally, to further optimize the recognition results, the generalization ability of the model was improved by adding spatial pyramid pooling and weight decay optimization, which optimizes the recognition results. The experimental results demonstrate that compared with other current depth models, this method can effectively locate ripe strawberries under complex background and is not easily affected by the interference environment. With higher recognition accuracy and sensitivity, in data set C, the recognition accuracy and sensitivity are the highest, reaching 92.46% and 94.28%, respectively.

Key words: strawberry recognition, depthwise separable convolution, residual network, deep learning, squeezeandexcitation blocks;data to enhance

摘要： 为解决自然状态下成熟草莓存在的背景干扰、信息丢失等问题，提出一种基于深度残差学习的草莓识别方法。首先，引入深度可分离卷积降低残差网络参数，从不同角度提取成熟草莓特征，通过交叉熵损失函数来识别分类层中的草莓。其次，嵌入压缩和激励模块学习特征权重，使用特征重新校准改善网络的学习和表征属性。最后，采用添加空间金字塔池化、加权衰减优化方法提高模型的泛化能力，优化识别结果。试验结果表明，和现有其他深度模型相比，该方法能够有效地定位复杂背景下的成熟草莓，不易受到干扰环境的影响，具有更高的识别准确率和灵敏度，在数据集C中的识别准确率和灵敏度最高，分别达到92.46%和94.28%。

关键词: 草莓识别, 深度可分离卷积, 残差网络, 深度学习, 压缩和激励模块, 数据增强