Research On Light-weight Convolutional Neural Object Detection Network

Object detection,one of the main research issues in computational vision,which incorporated technology such as feature extraction,machine learning,and deep learning,was widely used in production and living.Due to the strong feature expression ability and robustness,the object detection algorithms based on convolutional neural network were concerned.Because of the high computational complexity,too many weight parameters,and large size of detection models,it was difficult to be used in scenarios where computing power and memory were limited.To solve those problems the light-weight multi-object detection method IR-YOLO(Inverted Residual YOLO)was developed in this paper.Firstly,deep separable convolution was used to replace standard convolution to decouple the "spatial-cross-channel" feature learning method,thereby reducing parameter and calculation.Secondly,to reduce information loss and improve detection accuracy the batch normalization layer was introduced and the invested residual block was constructed based on depth separable convolutions.Then,the feature extraction network was composed of 6 down sampling layers and 6 invested residual blocks.In addition,referencing the feature pyramid idea,multi-scale feature map fusion learning method was adopted,that was,deep feature map was fused with shallow feature map through upsampling operations and finally two-scale feature maps were generated.Finally,a lightweight multi-object detection framework was built,which performed object detection on the generated two-scale feature maps.The IR-YOLO algorithm was experimentally verified.It compared with the pre-improved algorithm and the series of YOLO algorithm,and the algorithm performance was objectively evaluated in the aspects of calculation consumption,model size,detection speed and detection accuracy.The experimental results show that the light weight convolutional neural network object detection model in this paper can effectively reduce the amount of parameter and calculation on the basis of ensuring the detection efficiency,and thereby reducing the calculation complexity and model size of the model.And then the problem that the object detection model based on deep learning depended on the computing resources of the hardware platform is improved.There are 33 figures,17 tables and 62 references in this paper.