Localization and counting of string tomatoes based on improved Tiny-YOLOv5l algorithm

Abstract

Objective： This study is to improve the sorting efficiency of string tomatoes, and solve its false detection and false detection. Methods: First, collect the image data set of string tomatoes, expand the data and improve the generalization performance of the model through data enhancement, and 3×3 convolution is replaced by improved SVM-MHSA layer. By replacing softmax classification function in MHSA with SVM classification function which is more suitable for string tomatoes, the detection accuracy of string tomatoes is enhanced. Secondly, the remaining 3×3 convolution is replaced by deep separable convolution to reduce the number of parameters and improve the operation efficiency. Finally, random correction linear unit is introduced to improve the convergence speed of network training. Results: the test results show that the improved tiny YOLOv5l model can effectively realize the string single fruit recognition and positioning and the whole string fruit counting. The detection frame loss rate is reduced from 1.48% to 1.34%, the target loss rate is reduced from 1.98% to 1.73%, the confidence loss is reduced by 1.4%, the accuracy is increased from 97.36% to 98.89%, and the recall rate is increased from 97.35% to 98.56%. Conclusion: The tiny YOLOv5l algorithm is more accurate and lightweight. It has a high recognition accuracy in the face of challenges such as occlusion, background interference, illumination change and virtualization, and provides accurate information on the location of single fruit and the quantity of the whole string of fruit for post natal string tomato sorters.

Keywords

• deep learning
• classification function
• skewer tomatoes
• complex scenes