Image Understanding Lightweight Algorithm Design For Industrial Scenarios

Deep convolutional neural networks have been widely used in image classification,semantic segmentation,object detection and other fields,but traditional convolutional neural networks need to occupy a large amount of storage space while achieving high precision.The essence of lightweight neural networks lies in optimizing storage space and accelerating processing speed while maintaining the accuracy of deep neural networks.This type of network significantly reduces the requirements for storage capacity and computing power of devices,resulting in more efficient processing speeds,thus garnering widespread attention.For the design of lightweight networks,the core issue is how to achieve an optimal balance between algorithm efficiency(including space complexity and time complexity)and accuracy.This paper mainly focuses on the lightweight design of the image understanding network(i.e.,image recognition and semantic segmentation)in industrial scenarios.While reducing the network time and space complexity,it maintains a relatively excellent network representation ability,making the improved model maintain high precision in image understanding tasks such as image classification and semantic segmentation.Experiments in actual industrial scenarios show that the improved network meets the requirements of real-time performance and precision.The main work of this thesis is as follows:(1)The S-Ghost Bottleneck structure,a more efficient convolution module designed to reduce the computational cost of convolution operations and improve the feature representation ability of networks,has been developed for the application of graphite electrode label recognition in a carbon factory.Using this structure,a lightweight classification network LGSNet(Light Ghost Networks,LGSNet)was built,and quantitatively and qualitatively compared with multiple classification networks on different datasets.The experimental results show that the proposed network has significantly reduced parameter and computational costs without affecting accuracy and inference speed,and even improving them.The network has strong generalization ability,meets the requirements of the industrial scenario,and has good scalability.(2)In response to the challenges of sample collection complexity and loss of spatial information in real-time semantic segmentation tasks,this research aims to balance model inference speed,memory usage,and segmentation accuracy by applying the S-Ghost module to the feature extraction stage of semantic segmentation.Based on this,the researchers combined the spatial pyramid structure to obtain the real-time semantic segmentation network SGlraspp,the encoder-decoder architecture to obtain the real-time semantic segmentation network SGUnet,and the dual-path fusion module to obtain the real-time semantic segmentation network SGfusion.In order to verify the effectiveness of the designed architecture in the semantic segmentation task,comparative experiments are carried out on the DRIVE dataset and the graphite electrode dataset,and the network training loss function was improved to adapt to the special characteristics of these two datasets.From the perspective of model complexity,the three proposed models have significantly reduced time and space complexity,which helps to speed up the inference speed of the model and improve the real-time performance of the model;from the perspective of model performance,all three models can meet the requirements of real-time and accuracy.In terms of segmentation integrity and the processing of segmentation details,the fusion algorithm SGUnet of the S-Ghost module and the encoder-decoder architecture perform best.(3)The bottleneck module of S-Ghost is further optimized and the S-dual-Ghost module is proposed.Traditional convolution operations tend to obtain local receptive fields,and the resulting local features can easily lead to misclassification of objects and targets.In response to this problem,the S-Ghost module is combined with the self-attention dual mechanism to improve network representation performance: a positional attention mechanism is used to weight more extensive information into the local receptive field to enhance feature expression,and a channel-wise self-attention mechanism is used to improve the network’s ability to discriminate features.The experimental results show that the S-dual-Ghost fusion module reduces the network’s misclassification of the target and improves the network representational ability.To sum up,this thesis takes the design of an efficient convolution module as a starting point and conducts in-depth research on lightweight algorithms for industrial scene image understanding.As a result,high real-time lightweight image classification and semantic segmentation algorithms are proposed while maintaining comparable accuracy to similar algorithms.The algorithm’s considerable precision performance makes it capable of being applied in industrial scenarios.