Research on apple defect detection method based on improved CNN

Abstract

Objective: To solve the problems of low accuracy and poor efficiency in existing apple defect detection methods. Methods: Based on a fruit image acquisition system, an improved convolutional neural network was proposed for detecting surface defects in apples. Deep separable convolution was Introduced to replace the original network standard convolution, to improve the speed of feature extraction. The Leaky ReLU activation function was introduced to replace the ReLU activation function to improve the calculation efficiency and accuracy. Global average pooling was introduced to replace the fully connected layer, to reduce the computational complexity of the network model. After each layer of convolution, a batch normalization layer was added, and its superiority was verified through comparative analysis between experiments and conventional methods. Results: Compared with conventional methods, the proposed method had higher detection accuracy and speed in apple defect detection, and had fewer model parameters, with an accuracy rate of 99.60%, a detection speed of 526 FPS, and a model parameter quantity of 389 072. Conclusion: This apple defect detection method can effectively reduce model parameters and detection time, with high accuracy and speed.

Keywords

• defect detection
• apple
• convolutional neural network
• depth separable convolution
• leaky relu activation function
• global average pooling