Cotton Image Segmentation Network Based on Improved DeeplabV3+

Abstract

Aiming at the observation of cotton flow conditions in cotton production lines, a cotton image segmentation algorithm with improved DeeplabV3+ network is proposed, which introduces the lightweight network MobileNetV2 as the backbone feature extraction network; replaces the standard convolution in the void space pyramid pooling module with the depth separable convolution to compress the model size, and introduces the channel attention module to capture the image contextual information to effectively improve the segmentation accuracy of the model. The proposed algorithm achieves 96.86% pixel accuracy and 92.14% intersection ratio on the test set, which is 0.70% and 0.22% better than the original version, and the model parameter size is 15.29 MB, which is 92.7% smaller than the previous one, and the prediction time of a single frame is 18.67 ms, which is 65.8% smaller than the previous one. The experimental results show that the algorithm balances the characteristics of accuracy and real-time, and the overall comprehensive performance is optimal.