Optimized Object Detection Based on The Improved Lightweight Model Mini Net

Abstract

This paper proposes a Mini Net lightweight model that can be used for real-time detection. This model works together with Mini Lower and Mini Higher, which greatly improves the detection efficiency while ensuring the accuracy. The Mini module designs both the batch normalization layer and the excitation function at the front end of the module, which realizes efficient convolution, greatly reduces the amount of parameters and computation, and introduces the nonlinearity brought by more layers in the spatial dimension, which can improve the performance of the module extraction capacity. Based on the Mini convolution module, a multi-stage training strategy is proposed. The first stage makes the system fast and stable. In order to improve the overfitting phenomenon of the system, the second and third stages use finer features to improve the detection of small targets, thereby improving the Model training efficiency and detection accuracy.