| As one of the most promising branches of artificial intelligence,the deep learning method can effectively extract the abstract features of images.The object detection algorithms combined with deep learning have made a great breakthrough in the field of computer vision and achieve significant results.However,the current mainstream object detection altorithms have unsatisfactory detection results for small objects.The main reason is that the small objects have the problem of low resolution,low amount of information carried and susceptibility to background information interference,which leads to the loss of information of small objects when the network extracts features.Therefore,small object detection is one of the key challenges facing the current field of computer vision.Although the Path Aggregation Network(PAN)adopted by YOLOv5 can effectively alleviate the problem of semantic information loss for multi-scale objects,it still cannot fully recover the spatial information that has been lost for the objects.Thus,in response to the above problems,the YOLOv5 algorithm is used as a benchmark to be further researched in this thesis.The network structure of YOLOv5 is optimized for improving the sensitivity and detection effect for small objects.The specific work is as follows:Aiming at the small size of the objects,the Neck structure of YOLOv5 is taked as the starting point to study the impact of the structure of multi-scale feature fusion on the performance of small object detection.In order to retain more shallow details and location information for small objects,this thesis adds an additional layer of higher resolution(160× 160),which effectively utilizes the spatial information and semantic information of smaller objects and improves the model’s detection ability for small objects.To solve the problem of the loss of objects’ feature information caused by continuous convolution and downsampling of deep networks,this thesis conductd in-depth research and further improvement on the principle of the receptive field block.The receptive field of the feature maps was expanded to obtain the semantic information of small objects on the raw images,which enhanced the ability of the network for extracting the semantic information of objects in the shallow feature maps and reduced the missed detection of small objects in the algorithm.The E-RFs module proposed in this thesis was embedded in the Neck of the model to view the ability changes of feature enhancement of small objects.Aiming at the problem of complex information in the image’s background interfers with small objects,this thesis introduced the CBAM attention mechanism in the Neck of YOLOv5,which integrates channel and spatial information and enables the network to independently distinguish the relevance and effectiveness between different channels of feature maps.Additionly,the CBAM module is added after each concatenation module at the neck of the model.By introducing the attention mechanism in the feature fusion stage,the expression of deep feature information is enhanced to improve the ability of the network for detecting small objects.The above three methods are combined to reconstruct the original YOLOv5 for obtaining the proposed E-YOLOv5 model that contributed to detect small objects.Specifically,the multi-scale feature fusion layer is improved at first.The improved ERFs module is embedded in the appropriate location of the model after a higher resolution detection layer is added.And the CBAM attention module is introduced in the Neck part.Based on the Vis Drone-DET2018 dataset,this thesis conductes validation experiments on the E-YOLOv5 model.The experimental results show that the m AP value of the EYOLOv5 model increased by 7.05%,which verify that the proposed model in this thesis effectively enchances the ability of detecting small objects. |
