Research On Object Detection Method Based On MobileNet-SSD

Object detection is an important research field in computer vision and image processing.With the advent of the era of big data,the target detection method based on deep learning has become a research hotspot.Although the detection accuracy is far more than traditional methods,it still cannot meet Real-time and accuracy requirements.In order to balance the size and detection performance of the network model,this paper uses the lightweight network MobileNet for deep feature learning and combines with the SSD framework to achieve target detection.The formed MobileNet-SSD can effectively compress the size of the network model and improve detection rate.However,in the regression process,the Smooth L1 function does not reflect the overlap between the prediction box and the default box well,and there will be problems of missed detection and positioning deviation in the actual scene;in addition,the MobileNet-SSD network model uses feature pyramid for multi-scale target detection,but only Conv11 layer is used for the detection of small-ratio targets.Therefore,there are problems that the feature extraction is insufficient and the detection rate of the small-ratio target is low.In response to the above problems,the main innovations of the content of this study are as follows:(1)In view of the problem that the loss function in the MobileNet-SSD regression algorithm does not reflect the overlap between the prediction box and the default box,this paper proposes to use the improved cross merge ratio parameter Io U as the loss function to regress the target bounding box.Firstly to calculate the areas S and S'and the overlap area S~I of the prediction box B and the default box B'respectively;and then to find the smallest area C that contains the prediction box and the default box and calculate the area S~C of C to calculate the Io U of the prediction box and the default box;finally to use Io U minus the area of the area in C that does not cover the prediction box and the default box to obtain the improved Io U,than formed a new bounding box regression loss function.In the VOC2007 dataset,when the threshold is set to 0.5,the AP value increases by 0.024;when the threshold is set to 0.75,the AP value increases by 0.046.(2)Because of the problems,this paper firstly introduces Dilated Convolution after the Conv11 layer of the MobileNet-SSD network structure,setting the dilation rate to 4,maintaining the spatial resolution while improving the receptive field of the low-level feature map;afterwards in the feature mapping strategy,the image output by the model is used as the recommended region for default box mapping in order to enhance the ability to express low-level features.In the VOC 2007 dataset,the detection rate of the model in the image with higher target complexity can be increased to 81.1%,indicating that the detection accuracy of the target with a small proportion is improved,and the detection accuracy m AP can reach 81.3%,which effectively improve the detection performance of the model.