关键词: 多尺度卷积特征, SSD模型, MobileNet, 图像目标检测

Abstract: In order to improve the accuracy and efficiency of the SSD (single shot multibox detector) in the multi-task scene, which is a basic model of target detection in computer vision, according to the deep learning and relevant theoretical research, this paper constructs a multi-scale convolutional neural network structure with feature pyramid. It contains the basic ideas of a lightweight deep neural network called MobileNet. The following work is completed on Tensorflow platform. First, it amplifies the feature map of the low-layer convolution layer to reserve more target feature information, and then extracts the feature information of the high layer. Second, while filtering the overlap target candidate areas, it sets the threshold value to eliminate redundant target candidate areas based on the idea of non-maximum suppression. It reduces the number of negative samples and makes the model’s effect become stable gradually. Third, to ensure the stability of the model in this paper, it conducts the positive and negative samples in the target detection process of matching between the predicted region and the real region. Based on the research methods above, the recognition speed of the model in multi-target tasks has been accelerated, the robustness is better, the accuracy is higher, and the demands of the hardware configuration resources are reduced appropriately at the same time.

Key words: multi-scale convolution features, single shot multibox detector (SSD), MobileNet, image target detection