| Salient object detection is a crucial research direction in the domain of computer vision,which focuses on automatically identifying the most prominent regions or objects in an image.With the rapid advancement of information technology,the exponential growth of image data has elevated the importance of salient object detection in fundamental computer vision tasks.Such as image summarization,image search,and information discovery.Therefore,significant object detection has become one of the hot topics in the field of computer vision.The advent of deep learning technology has spurred a growing number of researchers to employ deep neural networks in salient object detection,leading to significant breakthroughs.However,there remain several challenges in this area.For example,the defects of single-feature detection and insufficient multi-feature fusion lead to high background noise and unclear edges in the significant map.Some methods currently available have a trade-off between accuracy and computational complexity,where some methods have high accuracy but require large computational resources,while others require less computational resources but have low accuracy.Therefore,how to reduce computational complexity while ensuring accuracy is an important issue.This paper mainly focuses on a lightweight significant object detection method based on convolutional neural networks,which implements multi-feature fusion.The specific research contents are as follows.(1)Targeting the drawbacks of current salient object detection methods,which include single feature detection and inadequate multi-feature fusion,a salient object detection method with multiscale visual perception and fusion is proposed.To begin with,multiscale visual perception module is constructed using atrous convolution to emulate the receptive fields in the visual cortex,while fully leveraging the potential of atrous convolution in convolutional neural networks.Global spatial information of significant objects is extracted step by step in the backbone network,effectively enhancing the foreground salient areas and suppressing the background noise areas.Then,a multiscale feature fusion module is designed to integrate highlevel semantic information with detail information by using feature pyramid and spatial attention mechanism,which can recover spatial structure information of salient objects more effectively while suppressing noise transmission.The two core modules make full use of multifeature fusion and the experimental results show that the edge contour of the salient graph detected by this method is clear,the background is clean,and the performance indexes are greatly improved on the benchmark network.(2)In response to the trade-off between detection accuracy and computational cost of current salient object detection methods,a lightweight salient object detection method based on multi-feature attention blocks is proposed by redesigning the network.The multi-feature attention block uses depth-wise separable convolution to extend the atrous convolution,and constructs a serial multi-feature fusion module based on the mechanism of human visual processing.While expanding the receptive field,multi-scale features are mined through adjacent progressive fusion.Then,channel attention and spatial attention are used to enhance feature extraction,and Ghost convolution’s inexpensive linear operation is used to avoid redundant feature maps and save computational cost.The feature extraction unit is reasonably constructed using basic blocks to build the backbone network encoder to extract features,followed by extracting context information through pyramid pooling modules.The decoder fuses global information,shallow features,and deep features to output the final saliency map.Experimental results show that,while maintaining 1.45 M parameters and 1.2G computational cost,the proposed lightweight model achieves detection performance comparable to some mainstream methods in terms of detection accuracy.(3)Academic practice is carried out on the above two theories.In view of the inefficiency and delay of manual detection of conveyor belt defects,a conveyor belt defects detection system based on significance detection method is proposed to realize the unmanned detection of conveyor belt defects and improve the detection efficiency.By using the lightweight significant object detection method,morphological filters of image processing and edge detection technology,the conveyor belt defect detection and location are realized.Based on python,the back end service of Web application is built by using Flask and the humanized UI of the front end is designed by using Vue,and the conveyor belt defect detection platform is realized.The test results demonstrate that the system satisfies the demands of defect detection in production environments. |
