Research On Lightweight Small Object Detection Model Based On SSD

With the development of convolutional neural networks,object detection based on deep learning have been widely used in the fields of autonomous driving,smart manufacturing,and smart home.However,convolutional neural networks often have high requirements on computing power and storage.Mainstream target detection models are difficult to achieve real-time detection effect on embedded devices with low computing power and low storage.At the same time,due to the deployment position and shooting angle of embedded devices in the real environment,the acquired image background is more complex and the object size is small,which brings greater challenges to the object detection model for embedded devices.For this,the dissertation studies the SSD(Single Shot Multi Box Detector)object detection model from three aspects of network structure,parameter quantization and feature fusion.It proposes two object detection models for embedded devices,as follows:(1)The dissertation proposes a real-time object detection model SSDLite-tiny suitable for embedded devices.First,the lightweight network Mobile Net V3 is used as the backbone network to extract image features so as to reduce the computational complexity and network parameters of the model.Then,feature fusion module is introduced in the detection part.The feature layer used for prediction in the model can obtain rich semantic information from the high level.Finally,the weight and input of the model are converted into INT8 type by perceptual quantization technology to further compress the model and accelerate the reasoning process.Using the Udacity dataset to evaluate SSDLite-tiny on the embedded device NVIDIA Jetson Xavier,its detection speed and the average accuracy of multiple categories(mean Average Precision,m AP)reached 75.08 frames per second and18.10% respectively.Combining different types of model compression algorithms in the above manner achieves the effect of balancing detection accuracy and detection speed.(2)In order to improve the detection accuracy of small object,the dissertation proposes a two-way feature fusion detection model SSDLite-attention based on the channel attention mechanism.First,the dissertation proposes a bidirectional feature enhancement module based on FPN,so as to make full use of the fine-grained information of shallow features.Then,the channel attention mechanism is used to mine the contextual semantic information of small object,and the shallow features and high-level features are deeply integrated,so as to improve the semantic representation ability of features to small object.In testing Udacity data sets on the embedded device NVIDIA Jetson Xavier,SSDLite-attention’s small object detection accuracy is 5.1% higher than SSDLite-tiny’s18.1%,and the detection speed is 31.6 frames per second.Experimental results show that bidirectional feature fusion guided by attentional mechanism can improve the accuracy of small object detection.Through sufficient comparative experiments,the two object detection models SSDLite-tiny and SSDLite-attention proposed in this dissertation based on SSD model can be successfully deployed on embedded devices and achieve real-time detection effect.Meanwhile,SSDLite-attention model can effectively improve the detection accuracy of small object.