In recent years, deep learning has been developed very quickly, and related research has shown a blossoming scene. Inception-v4 is a wide and deep network with good classification performance. The network structure is very complex, with convolution operations of different sizes, but there are two limitations: the inability to adaptively select the convolution kernel according to the characteristics of the image and the feature extraction from the high-level layer is not strong. This paper focuses on the investigation on the Inception-v4 model and has made several improvements. The improved Inception-v4 model is named BeIn-v4, which integrates the ideas of the Selective Kernel Network (SKNet) into the Inception-v4 network, and adjusts the network structure to achieve improvements. A number of comparative experiments have been carried out on the network before and after the improvements. The experimental results show that BeIn-v4 can obtain better classification results on the tested image datasets than Inception-v4.