A pepper quality detection and classification model based on transfer learning combined with convolutional neural network is proposed as a solution for low efficiency of manual pepper sorting at the current stage. The pepper dataset was amplified with data pre-processing methods including rotation, luminance switch, and contrast ratio switch. To improve training speed and precision, a network model was optimized with a fine-tuned VGG 16 model in this research, transfer learning was applied after parameter optimization, and comparative analysis was performed by combining ResNet50, MobileNet V2, and GoogLeNet models. It turned out that the VGG 16 model output anticipation precision was 98.14%, and the prediction loss rate was 0.0669 when the dropout was settled as 0.3, learning rate settled as 0.000001, batch normalization added, and ReLU as activation function. Comparing with other finetune models and network models, this model was of better anticipation performance, as well as faster and more stable convergence rate, which embodied the best performance. Considering the basis of transfer learning and integration with strong generalization and fitting capacity of the VGG 16 finetune model, it is feasible to apply this model to the external quality classification of pepper, thus offering technical reference for further realizing the automatic classification of pepper quality.

The efficient identification of the field flat jujube is the first condition to realize its automated picking. Consequently, a lightweight algorithm of target identification based on improved YOLOv5 (you only look once) is proposed to meet the requirements of high-accuracy and low-complexity. At first, the proposed method solves the imbalance of data distribution by improving the methods of data enhancement. Then, to improve the accuracy of the model, we adjust the structure and the number of the Concentrated-Comprehensive Convolution Block modules in the backbone network, and introduce the attention mechanisms of Efficient Channel Attention and Coordinate Attention. On this basis, this paper makes lightweight operations by using the Deep Separable Convolution to reduce the complexity of the model. Ultimately, the Complete Intersection over Union loss function and the non-maximum suppression of Distance Intersection over Union are used to optimize the loss function and the post-processing process, respectively. The experimental results show that the mean average precision of improved network reaches 97.4%, which increases by 1.7% compared with the original YOLOv5s network; and, the parameters, floating point of operations, and model size are compressed to 35.39%, 51.27%, and 37.5% of the original network, respectively. The comparison experiments are conducted around the proposed method and the common You Only Look Once target detection algorithms. The experimental results show that the mean average precision of the proposed method is 97.4%, which is higher than the 90.7%, 91.7%, and 88.4% of the YOLOv3, YOLOv4, and YOLOx-s algorithms, and the model size decreased to 2.3%, 2.2%, and 15.7%, respectively. The improved algorithm realizes a reduction of complexity and an increase in accuracy, it can be suitable for lightweight deployment to a mobile terminal at a later stage, and it provides a certain reference for the visual detection of picking robots.