SSD For Ochotona Curzoniae Object Detection Based On Multi-Scale Feature Fusion

The detection of Ochotona Curzoniae is essential for estimating population numbers and studying population dynamics.However,detecting Ochotona Curzoniae in natural scenes is challenging due to their small size,indistinct features,and complex backgrounds,resulting in few features for detection,which makes it difficult to detect.The SSD based on deep convolutional neural network utilizes the idea of an image pyramid network to detect objects of various scales effectively.However,the low-level features lack high-level semantic information leads to misclassification of small object,the accuracy of SSD in detecting small object is limited.To address this issue,this paper proposes a multi-scale feature fusion SSD small object detection model to improve the detection accuracy of small object in SSD based on deep CNN.Firstly,the multi-scale features extracted by the SSD backbone network are fused from top to bottom using an element-wise summation method based on the idea of the Feature Pyramid Network(FPN),to enrich the high-level semantic information of the low-level features.CBAM attention mechanism is also used to enhance the fused features.Secondly,the enhanced features are fused again using channel connections to achieve more sufficient feature fusion.Finally,the twice-fused features are down-sampled to generate a feature pyramid network,which is used to predict object and obtain the detection results.The detection results on Ochotona Curzoniae data set and NWPU VHR-10 data set show that the average accuracy of object detection model proposed in this paper is higher than other object detection models,and can improve the accuracy of small object detection.Although the multi-scale feature fusion SSD small object detection model based on FPN can improve the performance of SSD in object detection to some extent,it does not consider the significant semantic differences between different scales of features.Fusing multiple feature layers with significant semantic gaps can decrease the multiscale feature representation ability.To address this issue,this paper introduces a deep supervision mechanism to the multi-scale feature fusion SSD small object detection model.A deep supervision method is used to add the same supervisory signal to each feature layer before the top-down feature fusion,allowing different feature layers to learn similar semantic information under the same supervisory signal to narrow the semantic gap between feature layers and improve the multi-scale feature representation ability,thereby enhancing the detection accuracy of small object.The detection results on the Ochotona Curzoniae dataset and the NWPU VHR-10 dataset show that the proposed detection model outperforms other detection models in terms of average detection accuracy and can effectively improves the accuracy of object detection model for small object.