Design of vision recognition system for picking robots

Abstract

Based on yolov4-tiny deep learning neural network, an improved yolov4-tiny network model is proposed in order to achieve the reduction of the network model for overlapping fruits and branch-obscured fruits in natural environment and to realize the accurate and fast recognition of apple-pear fruits, the main improvement measures include: firstly, the CSPBlock residual network module of the backbone network is introduced in the module of the backbone network to replace the 3×3 convolution kernel in it, which improves the perceptual field of the feature layer in the network and enhances the extraction capability of the target feature information through the spatial consistency and channel specificity of the Involution operator. second, the output of the first layer of the CSPBlock module in the backbone network containing the rich surface information of the image is extracted in the feature pyramid with the first and second scale feature maps for multi-scale feature fusion to enhance the extraction capability of dense small target feature information, by conducting training experiments on the apple pear dataset collected by ourselves, the experimental results show that the accuracy of the improved network structure is 95.45%, an improvement of 2.84%, and the recall rate is 94.92%, an improvement of 2.83%, compared with yolov4-tiny, the improved method improves the accuracy of fruit recognition and provides the theoretical basis for the subsequent apple pear picking robot to quickly identify picked apple-pears.